ResQ-UAV: A Novel Dataset Supporting Robust Recognition for Future Drone Rescue Missions

ResQ-UAV: A Novel Dataset Supporting Robust Recognition for Future Drone Rescue Missions | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article ResQ-UAV: A Novel Dataset Supporting Robust Recognition for Future Drone Rescue Missions Jian Deng, Honghai Zhang, Zeyu Liu, Zihan Yu This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8741863/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract This paper presents and open-sources ResQ-UAV, a dataset designed for person detection in corrupted videos from the perspective of Unmanned Aerial Vehicles (UAVs). Generated using H.264 bitstream corruption techniques, the dataset aims to simulate realistic non-linear signal impairments encountered in actual image transmission links. By mimicking authentic communication link failures, the images exhibit highly complex degradation characteristics, posing significant challenges for detection algorithms operating under communication-constrained conditions. ResQ-UAV comprises 16,414 frames derived from 32 video sequences, featuring a total of 282,095 meticulously annotated bounding boxes for persons. It encompasses diverse complex scenarios, including urban arterials and suburban areas, across various lighting conditions (day and night). Serving as a dedicated benchmark for evaluating video transmission robustness, this dataset significantly enriches the data resources available for UAV vision in complex transmission environments. Benchmarking results based on various state-of-the-art object detection algorithms demonstrate that ResQ-UAV establishes a rigorous performance baseline for multi-object detection tasks within corrupted video environments. The dataset is poised to provide a reliable data foundation and verification platform for the research and application of critical technologies in disaster rescue operations. Data and code are available at: https://github.com/dj-dengjian/ResQ-UAV.git . Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Background & Summary In recent years, the frequent occurrence of natural disasters globally—such as floods, earthquakes, and typhoons 1 —has posed significant threats to human life and property. During the critical "golden rescue window" following such catastrophes, the complex and dynamic post-disaster environment 2 often leads to the paralysis of ground transportation, communication interruptions, and risks of secondary disasters, thereby imposing severe impediments to traditional manual search and rescue operations. Against this backdrop, Unmanned Aerial Vehicles (UAVs) 3 , leveraging their high mobility, flexible deployment, and extensive aerial field of view, have demonstrated broad application prospects in post-disaster emergency response and humanitarian aid. Particularly in Search and Rescue (SAR) missions, UAVs 4 can rapidly penetrate areas that are inaccessible or hazardous to humans. They facilitate high-efficiency, all-weather scanning of large-scale regions, thereby enabling the rapid localization of trapped individuals and significantly enhancing rescue response speed and success rates. To address the specific challenges of UAV visual perception, the academic community has recently introduced several targeted, high-quality datasets. Addressing the challenge of "small objects and high density," Manipal-UAV (2023) 5 provides large-scale annotated data containing over 150,000 instances, specifically designed to evaluate model detection capabilities under extreme scale variations. Furthermore, Unicamp-UAV (2026) 6 incorporates a negative sample mechanism to focus on the issue of high false-positive rates in open environments, representing the latest standards in data acquisition and annotation. Regarding "specific rescue scenarios," the Stanford Drone Dataset (2020) 7 offers trajectory analysis data from a top-down aerial perspective, while the recent NITC (2025) 8 focuses on person detection against building backgrounds, aiming to resolve recognition difficulties caused by occlusion from structural elements. Although existing datasets have enriched UAV vision research, the field still faces severe challenges when applied to realistic post-disaster rescue scenarios, primarily manifested in the following aspects: Scarcity of data with video transmission link impairments : Existing UAV datasets typically focus on environmental challenges at the optical acquisition end, such as fog, rain, or low-light conditions. However, in actual Beyond Visual Line of Sight (BVLOS) rescue or logistics missions, limited by bandwidth fluctuations and electromagnetic interference, the image transmission link often suffers from specific bitstream corruptions (e.g., mosaic effects and frame loss caused by H.264/H.265 decoding errors 9 ). Most currently public datasets consist of locally stored lossless images, lacking corrupted samples that reflect real-world communication link failures. Consequently, models trained on existing data are prone to generalization failure when encountering real-time transmission artifacts, leading to severe missed detections. Discrepancy between the ideal training domain and the real-world deployment domain : The performance of deep learning models relies heavily on the consistency between training and testing data distributions. Existing benchmark models are typically trained on clear, high-quality "source domain" data, assuming that test scenarios contain only simple linear noise (e.g., Gaussian blur). However, real-world video transmission impairments exhibit highly non-linear and structured characteristics (such as blocking artifacts), constituting a significant "domain shift." Due to the lack of adversarial training data specifically targeting transmission-corrupted scenarios, existing research struggles to effectively evaluate and improve model robustness under this specific domain discrepancy. To bridge the gap between optical imaging and video transmission stability assessment, we constructed and open-sourced the ResQ-UAV dataset, designed for future rescue research (as shown in Fig. 1 ). Unlike traditional acquisition methods, the core construction strategy of ResQ-UAV involves processing existing high-quality UAV data by simulating authentic H.264 video coding transmission errors. This process generates corrupted samples possessing typical degradation characteristics associated with bitstream corruption. As illustrated in Fig. 1 , the ResQ-UAV dataset provides a visual representation of bitstream corruption degradation characteristics across diverse scene perspectives. These images faithfully reproduce the visual impairments commonly encountered in communication-constrained disaster sites, including non-linear noise such as severe blocking artifacts, color distortion, local pixel loss, and image tearing. The visualization results in Fig. 1 clearly demonstrate how such structured degradation blurs the edge features of person targets, leading to indistinguishability between targets and the background. Consequently, this provides a challenging and indispensable foundational data resource for developing robust recognition technologies tailored to complex disaster environments. The ResQ-UAV dataset comprises 16,414 frames derived from 32 video sequences, spanning a spectrum of transmission states from lossless to severely corrupted. The dataset contains a total of 282,095 meticulously cleaned bounding boxes for persons. The scenarios encompass a variety of complex environments, including urban arterials, intersections, suburban areas, and playgrounds, covering both diurnal and nocturnal periods. To facilitate research on domain adaptation algorithms, we constructed a paired "clean-corrupted" data structure and performed semantic merging and unified annotation for the "pedestrian" and "person" categories within the original data. To assess the challenging nature of the ResQ-UAV dataset in the field of robustness detection, we trained and evaluated four representative state-of-the-art (SOTA) object detection algorithms: the classic Faster R-CNN 10 , the multi-stage Cascade R-CNN 11 , the latest real-time detector RTMDet 12 , and the Transformer-based DINO 13 . Experimental results indicate that, compared to their performance on clean datasets, existing SOTA methods exhibit a marked degradation in performance (a substantial attenuation in mAP) on ResQ-UAV. This decline is primarily attributed to the prevalent feature loss caused by bitstream corruption within ResQ-UAV, demonstrating that current detectors lack the feature extraction capabilities and interference resilience necessary to handle such structured noise. As an exploratory initiative oriented towards disaster rescue scenarios, ResQ-UAV establishes a UAV person detection benchmark focused on evaluating video transmission robustness. By simulating bitstream corruption characteristics inherent to real-world communication-constrained environments, this dataset augments the sample diversity of existing UAV vision research under complex transmission conditions. It not only provides essential test data for research on small person detection in low-bitrate and high-packet-loss environments but also serves as a critical data foundation and validation platform for developing UAV Search and Rescue (SAR) 14 vision systems capable of adapting to complex channel interference with enhanced transmission robustness. Materials Date sources To construct a dataset of corrupted videos with high realism and significant challenge, this study selected the VisDrone2019-VID 15 dataset as the raw video benchmark. VisDrone stands as one of the largest and most authoritative benchmarks in the field of UAV visual analysis. The rationale for selecting it as the foundation for ResQ-UAV is primarily based on the following three considerations: First, the temporal continuity of video sequences. Unlike datasets based on static images such as DOTA 16 or TinyPerson 17 , VisDrone-VID provides complete video sequences. This characteristic is critical for simulating H.264 bitstream corruption. Since video coding relies on inter-frame prediction and Group of Pictures (GOP) structures 18 , single static images cannot capture the temporal propagation effects of decoding errors. The continuous frame structure of VisDrone enables us to authentically reproduce the non-linear dynamic degradation process where "corruption in a preceding frame leads to artifacts in subsequent frames." Second, extremely rich scene diversity. The original dataset was collected from 14 different cities across China 15 , spanning thousands of kilometers geographically. It covers a wide range of environments, including urban arterials, busy intersections, quiet suburbs, playgrounds, and residential areas. This extensive coverage ensures the complexity of background textures, effectively evaluating model robustness when background clutter is superimposed with corruption artifacts, thereby preventing the model from overfitting to a single background type. Finally, authentic UAV perspective challenges. The data acquisition encompasses various lighting conditions (daytime, nighttime, overcast) as well as different shooting angles (top-down, oblique) and flight altitudes. This implies that the original videos inherently contain UAV-specific challenges such as motion blur and drastic scale variations. Superimposing bitstream corruption on top of these inherent optical challenges allows for the construction of a "double difficulty" test benchmark that most closely approximates real-world post-disaster search and rescue scenarios. In summary, the video stream characteristics and high scene complexity of VisDrone-VID make it the ideal substrate for generating bitstream corruption benchmarks. Consequently, this study selected 32 video clips containing abundant human activities from the dataset as the core data source. Data processing Selection of Corruption Scheme In the construction of a corrupted video dataset, the selection of the data generation scheme is critical, as it directly determines the realism and challenge of the evaluation benchmark. Existing research on video inpainting and error concealment 19 predominantly adopts traditional "mask simulation" strategies, which involve directly overlaying error masks of predefined shapes onto decoded video frames. These masks typically manifest as regular slices or rectangular blocks, with positions and dimensions that are often fixed or follow simple random distributions. However, such pixel-level simulation methods possess notable limitations, failing to capture the physical characteristics of video data loss inherent in real-world multimedia communications. In actual UAV image transmission links, video streams undergo compression via standards such as H.264/AVC. Consequently, packet loss or signal interference during transmission occurs within the underlying "bitstream domain" rather than the "pixel domain" 20 . Governed by the inherent inter-frame prediction mechanisms and Context-Adaptive Variable Length Coding (VLC/CABAC) characteristics of H.264 encoding, minor perturbations in the bitstream layer induce highly complex visual artifacts at the decoder, exhibiting non-linear propagation characteristics. In contrast to simple artificial masks, errors precipitated by bitstream corruption are more diverse in typology. They encompass severe blocking artifacts, color artifacts 21 , texture loss, and global or local frame misalignment resulting from damaged motion vectors. Therefore, to bridge the gap between laboratory simulations and real-world operational failures, this study adopts a bitstream-based corruption scheme to generate corrupted samples that possess both high realism and stochasticity. Bitstream Corruption Model To generate corrupted data that is both controllable and possesses high realism, this study develops a bitstream corruption methodology based on the underlying syntax of video coding. Distinct from traditional pixel-level masking, this method adheres to the error control mechanisms of the H.264/AVC standard within IP network transmission 22 , constructing a universal framework capable of precisely simulating real-world network packet loss and bit errors. Specifically, leveraging the Network Abstraction Layer (NAL) structure and error propagation characteristics of H.264 23 , we propose a three-parameter stochastic corruption model (P, L, S), utilizing this tuple to precisely define the patterns and intensity of the degradation: Frame Corruption Probability (P) : This parameter simulates the packet loss rate in unreliable channels and dictates the frequency of corrupted frames within the video sequence, representing the statistical proportion of target frames randomly selected for corruption within a Group of Pictures (GOP). Corruption Start Position (L) : Refers to the byte offset within the payload of the selected Network Abstraction Layer Unit (NALU) where error injection or data truncation occurs. Corrupted Segment Length (S) : Defines the length of the continuous bitstream data (in bytes) to be removed or tampered with, used to simulate varying degrees of burst errors. The implementation logic of this method is strictly grounded in the parsing of the H.264/AVC bitstream syntax structure 9 . A standard H.264 bitstream is composed of a series of NALUs, wherein key control information—such as the Sequence Parameter Set (SPS), Picture Parameter Set (PPS), and Slice Header—constitutes a negligible proportion (< 0.01%) of the total bitstream but is vital for decoding synchronization. To prevent decoder crashes or catastrophic sequence disruptions, the model adopts an Unequal Error Protection (UEP) strategy. This involves bypassing high-sensitivity header information and targeting only the dominant Video Coding Layer (VCL) data for specific segment removal operations. Through this low-level manipulation, the generated corrupted video not only retains a valid syntax structure but also faithfully replicates a variety of highly non-linear, realistic visual impairments, including blocking artifacts, texture loss, and motion vector misalignment. Parameter Configuration and Adaptive Adjustment Although this study adopts a general bitstream corruption framework, given the substantial disparity in spatial scale between ResQ-UAV (derived from VisDrone) and traditional low-resolution datasets, we have performed a targeted adaptive reconfiguration of the corruption parameters by integrating video transmission characteristics with visual perception theory: Corruption Probability (P): Realistic Fitting of Communication Links. In this study, the frame corruption probability is set to P = 1/16, implying that, on average, one frame within a Group of Pictures (GOP) of length 16 undergoes bitstream corruption. This threshold setting references the realistic fading and packet loss models of UAV Air-to-Ground (A2G) propagation channels 24 . This ratio constitutes a critical equilibrium point: on one hand, it is sufficient to simulate significant visual interference generated under unreliable channel conditions; on the other hand, it adheres to the stability constraints of streaming media transmission, avoiding decoder resets or complete video stream interruptions caused by excessive error rates (Packet Loss Rate > 10%). Consequently, this accurately reflects the characteristics of real-world scenarios where UAVs experience intermittent packet loss in weak network environments. Corruption Position (L): Intensifying Cascaded Error Propagation. To maximize the attack efficacy of a single corruption event, we set the corruption start position to L = 0.2, meaning that error injection is primarily concentrated within the first 20% of the NALU payload. According to quantitative analyses of the impact of packet loss on subjective video quality 25 , the initial segment of a NALU typically carries critical decoding syntax elements such as Motion Vector Differences (MVD) and Macroblock Modes (MbMode). Corruption at an early position not only directly destroys subsequent data within the current frame but also leverages the inter-frame prediction mechanism of H.264 to trigger a severe "avalanche effect," causing temporal error diffusion to subsequent dependent frames (P-frames/B-frames). This strategic position selection significantly enhances the persistence and destructive power of artifacts, thereby substantially increasing the difficulty of robustness testing for detection algorithms. Corruption Size (S): Resolution-Based Perceptual Scale Augmentation. The parameter augmentation targeting high-resolution video represents the most significant adjustment in this study. Existing research on corrupted video is predominantly based on low-resolution materials (480P or 720P), where the commonly used corruption size of 4096 bytes produces only negligible visual noise in the 4K (3840×2160) Ultra-High-Definition (UHD) bitstreams widely adopted in the VisDrone dataset, failing to constitute effective interference. According to multi-scale perception theory in Spatio-Temporal Video Quality Assessment (ST-VQA) 26 , the saliency of visual artifacts is highly dependent on their proportion relative to the overall image. Therefore, to ensure that the corruption effects maintain strong visual impact and detection challenges within 4K frames, we substantially increased the corruption size to S = 20,480 bytes. This magnitude adjustment ensures that the blocking artifacts and signal tearing in ResQ-UAV match the scale of high-resolution UAV imagery, thereby enabling effective evaluation of model generalization capabilities under extreme conditions characterized by "large field of view, small objects, and strong interference." Object annotation To construct a benchmark suitable for high-precision personnel search and rescue as well as surveillance assessment, this study implemented a rigorous three-stage data processing pipeline consisting of "Reconstruction–Mapping–Cleaning." This workflow aims to resolve the semantic ambiguity inherent in the original labels and ensure the spatial precision and visual validity of the annotated data within the Corrupted Domain. Semantic Reconstruction and Full Re-annotation of Source Domain Data . Although the original VisDrone dataset provides rich annotations, there exists significant overlap and ambiguity in the semantic definitions between the "pedestrian" and "people" categories. To eliminate the potential interference of label noise on subsequent model evaluation, this study discarded the original label system, referencing consistent annotation standards used in wild crowded pedestrian detection 27 . Utilizing the intelligent annotation tool X-AnyLabeling, we conducted a comprehensive manual re-annotation from scratch on the 32 selected high-definition raw video sequences. In this process, all human targets were uniformly categorized into a single class, "Person," thereby ensuring high semantic consistency and purity of the Ground Truth data in the source domain. Automatic Projection Mapping of Cross-Domain Labels. Upon establishing high-quality annotations for the source domain, we developed a semi-automated transfer workflow from the original domain to the corrupted domain. Adhering to the principle of geometric invariance in synthetic data domain adaptation research 28 —which posits that H.264 bitstream corruption affects only image texture without altering the physical spatial coordinates of objects—we directly projected and mapped the re-annotated bounding boxes from the source domain onto the corresponding ResQ-UAV corrupted video sequences. This strategy guarantees pixel-level alignment of spatial positions between Clean Samples and Corrupted Samples, providing a perfect comparative baseline for subsequently quantifying the performance degradation of algorithms under identical scenes but varying image quality conditions. Cleaning Verification Based on Visual Perceptibility. Mere coordinate mapping may retain targets that are visually completely lost, thereby violating the general criterion of "Visual Perceptibility" in dataset construction 29 . Therefore, following the completion of automatic mapping, we introduced a rigorous manual cleaning phase. Annotators used X-AnyLabeling to perform frame-by-frame verification on the corrupted videos, focusing on the removal of targets that became completely unrecognizable to the naked eye due to frame loss, extensive mosaic occlusion, or severe color distortion. Through this physical filtering mechanism, the final annotation set generated for ResQ-UAV retains precise spatial localization while effectively excluding visually imperceptible invalid samples, thus faithfully reflecting the effective boundaries of detection algorithms in extreme visual environments. Data Records The ResQ-UAV dataset is hosted on a GitHub repository ( https://github.com/dj-dengjian/ResQ-UAV.git ), providing open access to the research community. Designed to maximize usability and support paired experimental analysis, the dataset features a highly structured hierarchical directory organization. The detailed file storage architecture and branching logic are depicted in Fig. 2 . The root directory of the dataset comprises three core components: the data folder for storing image sequences, the annotations folder for storing detection labels, and the mmdetection codebase integrated with benchmark detection algorithms. Detailed descriptions of each component are provided below: Image Data Organization (data). The data directory adopts a strict dual-branch parallel architecture to ensure a one-to-one correspondence between the clean source domain and the corrupted target domain: gt branch : Stores high-resolution lossless video frames extracted from the original VisDrone dataset, serving as the reference benchmark for visual perception. corrupted branch : Stores corrupted video frames generated based on the H.264 bitstream corruption model, serving as the target data for robustness evaluation. These two branches share an identical subdirectory structure, both indexed by video sequence ID. Regarding data partitioning, the 32 video sequences are strictly divided into three mutually exclusive subsets: train (containing 22 sequences for model training), val (containing 4 sequences for hyperparameter validation), and test (containing 6 sequences for final performance evaluation). All images are stored in JPG format and are accompanied by frame-level metadata for traceability. Annotation Data Organization (annotations). To accommodate the requirements of supervised learning and evaluation within corrupted environments, the annotations directory provides two sets of tailored ground truth labels, both organized in the standard MS COCO JSON format 30 : annotations_gt : Contains high-quality ground truth labels for the "Person" category following the comprehensive re-annotation of pristine original images. These are utilized for training "upper bound" models or conducting source domain supervision. annotations_corrupted : Contains ground truth labels for corrupted images processed through the "projection mapping–manual cleaning" pipeline. This annotation set excludes targets that are visually indiscernible within severely corrupted fields of view, thereby ensuring the objectivity of the evaluation. To facilitate experimental configuration, annotation information for all subsets (Train/Val/Test) has been merged into independent JSON files (e.g., train_merged.json). Researchers can seamlessly transition between "clean supervision" and "corrupted supervision" experimental settings by simply modifying the file paths in the configuration file. Technical Validation Model Environment All experiments were conducted on a high-performance computing workstation equipped with an 8-core processor, 32GB of RAM, and an NVIDIA Tesla V100 GPU with 16GB of video memory. The software environment was built upon Python 3.8, utilizing the open-source MMDetection framework 31 for the entire pipeline of model training and inference. To comprehensively evaluate the robustness of detection architectures within corrupted video environments, this study selected four representative algorithms encompassing distinct paradigms: the classic two-stage detector Faster R-CNN based on ResNet-50; the multi-stage cascade architecture Cascade R-CNN, focused on enhancing detection quality at high IoU thresholds; the high-performance one-stage anchor-free detector RTMDet integrated with a CSPNeXt backbone; and the Transformer-based 32 end-to-end detection paradigm DINO. Regarding training strategies, all models were initialized with official pre-trained weights from the MS COCO 2017 dataset to accelerate convergence and enhance the generalization capability of feature extraction. Tailored to the optimization characteristics of different architectures, the experiments employed either SGD 33 or AdamW 34 optimizers, while uniformly adopting the Cosine Annealing algorithm 35 for learning rate decay. This strategy, through periodic warm restarts and a smooth descent mechanism, effectively assists the model in circumventing local optima within complex loss surfaces, thereby facilitating convergence towards the global optimal solution. Evaluation Metrics To comprehensively assess the holistic performance of models under UAV perspectives and bitstream-corrupted environments, this study constructs an integrated evaluation framework encompassing detection accuracy, training dynamics, and computational efficiency. Detection Accuracy and Multi-Scale Perception Metrics. We strictly adhere to the evaluation protocols of the MS COCO object detection benchmark, adopting Average Precision (AP) as the core evaluation metric. Specifically, we report AP, defined as the mean Average Precision calculated over Intersection over Union (IoU) thresholds from 0.50 to 0.95 (with a step size of 0.05). This metric reflects the comprehensive performance of the model under varying degrees of localization rigor and serves as the primary criterion for measuring detector robustness. Furthermore, given the predominance of small objects in UAV datasets and their susceptibility to loss following bitstream corruption, we introduce fine-grained scale-specific evaluation metrics: AP50 : The average precision calculated at a fixed IoU threshold of 0.50. In the field of UAV vision, due to the extremely small size of targets and complex backgrounds, high IoU matching is often excessively stringent. Therefore, AP50} is widely regarded as a critical benchmark metric for assessing Small Object Detection performance. APs : The average precision specifically calculated for small objects (pixel area < 32²). In corrupted video detection tasks, the fluctuation of APs is particularly critical because blocking artifacts and ringing artifacts introduced by H.264 encoding often submerge the geometric features of tiny objects first, leading to detection failure. Training Dynamics and Stability Analysis. To verify the model's fitting degree and generalization boundaries regarding corrupted data, we conduct an in-depth analysis of convergence behavior during training. By visualizing the evolution trends of Classification Loss, Regression Loss, and mAP across both training and validation sets throughout the iteration process, we can effectively monitor risks of overfitting or underfitting. Particularly when comparing Transformer architectures (e.g., DINO) with CNN architectures (e.g., Faster R-CNN), analyzing differences in their convergence speeds and final steady states helps reveal disparities in learning efficiency and robustness regarding different inductive biases when processing non-linear signal noise. Computational Complexity and Deployment Feasibility. Given that UAV onboard computing platforms are typically limited by strict Size, Weight, and Power (SWaP) constraints, the spatiotemporal complexity of the model is key to determining the feasibility of its engineering implementation. Consequently, we introduce Parameters and Floating Point Operations (FLOPs) to quantify the model's static spatial footprint and dynamic computational load. Simultaneously, to assess real-time processing capabilities, we test the single-frame inference Latency and Frames Per Second (FPS) of each model under a unified hardware environment. By constructing the Pareto Frontier 36 of accuracy (mAP) versus speed (FPS), we aim to identify the optimal equilibrium point that satisfies real-time response requirements while ensuring high-precision detection. Finally, to quantify the specific impact of bitstream corruption on the detection system, beyond directly observing absolute accuracy on the corrupted dataset, this study introduces the Relative Performance Drop (RPD) as a robustness evaluation metric 37 . Drawing from methods in image quality assessment regarding the resilience of deep networks against distortion, this metric intuitively reflects the model's resilience against interference during video transmission by calculating the relative magnitude of change in evaluation values between the original validation set (Clean) and the corrupted validation set (Corrupted). A lower RPD value indicates superior performance retention during image quality degradation, suggesting that the feature extraction network possesses stronger robustness against non-natural artifacts. This is of significant importance for evaluating algorithm reliability in extreme environments such as post-disaster search and rescue. The RPD formula is defined as follows: $${\text{RPD}}=({\text{A}}{{\text{P}}_{{\text{gt}}}} - {\text{A}}{{\text{P}}_{{\text{corrupted}}}})/{\text{A}}{{\text{P}}_{{\text{gt}}}}$$ 1 To evaluate the performance recovery of the model following retraining with the introduced ResQ-UAV dataset, we propose the Performance Gain (GAP) metric. GAP is defined as the difference between the precision achieved by the model after training on corrupted data ( \({\text{A}}{{\text{P}}_{{\text{retrained}}}}\) ) and the baseline precision obtained when trained exclusively on clean data ( \({\text{A}}{{\text{P}}_{{\text{corrupted}}}}\) ). A positive GAP value directly quantifies the effectiveness of the proposed dataset in mitigating Domain Shift and enhancing the detector's adaptability to corrupted signals, serving as a pivotal criterion for validating the core contribution of this study. The formula for GAP is defined as follows: $${\text{GAP}}={\text{A}}{{\text{P}}_{{\text{retrained}}}} - {\text{A}}{{\text{P}}_{{\text{corrupted}}}}$$ 2 Analysis of Results Robustness Analysis of Existing Models To assess the generalization capability of mainstream object detection algorithms within video transmission-corrupted scenarios, we established an evaluation benchmark based on "source-domain supervision." In this experiment, four representative models—RTMDet, Faster R-CNN, Cascade R-CNN, and DINO—underwent parameter optimization utilizing exclusively the original VisDrone training set. Subsequently, inference evaluations were executed on both the clean original test set (GT Test) and the ResQ-UAV corrupted test set (Corrupted Test) constructed in this study. Table 1 details the performance comparison of each model across the two domains; results indicate that all algorithms exhibited varying degrees of accuracy degradation within the corrupted environment. To further investigate the intrinsic mechanisms underlying this cross-domain performance disparity, we employed the RTMDet model as a case study to conduct a visual analysis of its detailed training process and statistical characteristics under both the clean source domain (GT) and the corrupted target domain (Corrupted), as illustrated in Fig. 3 . Specifically, sub-figure (a) depicts the trajectory of training accuracy improvement as iteration epochs increase. It lucidly reveals that when training on corrupted data, the ascent rate of key metrics such as $ mAP $ and $ AP_{50} $ lags behind, and the ultimate peak is significantly lower than the clean baseline. This is further corroborated by the loss convergence curves in sub-figure (c), which demonstrate that the model convergence process within the corrupted environment is accompanied by higher loss values and greater oscillation amplitude. Concurrently, sub-figures (b) and (d) utilize violin plots to quantify the probability distribution characteristics of performance metrics and loss values. The results indicate that the $ mAP $ distribution for corrupted data is flatter with higher dispersion, while the loss distribution exhibits a pronounced long-tail effect. This disparity in distributional characteristics implies that the substantial volume of "hard samples" generated by bitstream corruption significantly exacerbates the optimization difficulty and uncertainty of the model within the feature space, thereby substantiating the fundamental cause of the quantitative performance degradation observed in Table 1 . Figure 4 further visualizes the differences in training dynamics of the classic two-stage detector, Faster R-CNN, between the clean source domain (GT) and the corrupted target domain (Corrupted). Observing the accuracy evolution trajectory in sub-figure (a), it is evident that although the model exhibits high stability throughout the training cycle (characterized by relatively smooth curves), the mAP metrics on the corrupted data (blue-toned curves) are consistently suppressed below the clean baseline (red-toned curves), forming a significant and persistent "performance gap." This phenomenon is corroborated by the loss convergence curves in sub-figure (c); while the model is capable of converging within the corrupted environment, its loss baseline remains persistently higher than that of the clean environment, indicating that video artifacts continuously interfere with the feature matching process of the Region Proposal Network (RPN). Furthermore, the violin plots in sub-figure (b) show that the mAP distribution of Faster R-CNN is extremely compact, reflecting the low-variance characteristic of its training process; however, the center of mass of the corrupted group's distribution is shifted downward overall. Meanwhile, sub-figure (d) reveals that the loss distribution of the corrupted data possesses a longer upper tail, suggesting the presence of a substantial number of hard samples within corrupted video frames that induce high regression errors, thereby limiting the model's ultimate performance ceiling in extreme environments. Figure 5 focuses on the training characteristics of the multi-stage detection architecture, Cascade R-CNN, across different data domains. By analyzing the accuracy evolution trajectory in sub-figure (a), it is evident that although the Cascade architecture is designed to progressively enhance detection quality at high IoU thresholds via a cascade regression mechanism, the upward trend of its mAP metrics on corrupted data (blue-toned curves) significantly lags behind the clean baseline (red-toned curves). Furthermore, the performance gap between the two domains does not notably narrow as training epochs increase. This phenomenon is further corroborated by the loss convergence curves in sub-figure (c), where loss values in the corrupted environment consistently remain at a higher level. This indicates that the non-linear noise introduced by video transmission severely hinders the refined feature alignment performed by the multi-stage detection heads. Additionally, the statistical distribution plots in sub-figures (b) and (d) demonstrate that, compared to the highly concentrated performance distribution observed with clean data, the corrupted data causes an overall downward shift and increased dispersion in the mAP distribution. Simultaneously, the long-tail distribution characteristics of the loss values suggest that, even with the adoption of a cascade optimization strategy, the model continues to face significant uncertainty and optimization difficulties when confronting severe blocking artifacts and blurring artifacts. Figure 6 finally illustrates the distinct training behavior of DINO, a Transformer-based end-to-end detection paradigm, when confronted with video corruption. Analysis of the accuracy evolution trajectory in sub-figure (a) indicates that, despite DINO possessing robust global context awareness capability via its self-attention mechanism, the ascent rate and final convergence value of its mAP metrics on corrupted data (blue-toned curves) remain significantly lower than the clean baseline (red-toned curves). This characteristic is manifested uniquely in the loss convergence curves in sub-figure (c). Unlike CNN architectures, DINO's loss descent curves are extremely smooth, suggesting a relatively stable optimization process; however, a constant "parallel gap" persists between the clean and corrupted curves. This implies that bitstream noise induces a systematic bias in the Bipartite Matching process. Furthermore, comparing the loss distributions in sub-figure (d) reveals that the loss values for clean data exhibit an extremely sharp and concentrated "needle-like" distribution, reflecting the Transformer's exceptionally high fitting certainty regarding clear samples. However, within the corrupted environment, the base of this distribution significantly broadens and exhibits a trend toward dispersion. This indicates that while the model is capable of processing the majority of corrupted samples, video artifacts severely impair the feature semantics of certain targets. Consequently, the model fails to effectively focus when processing these long-tail "hard samples," thereby limiting any breakthrough in overall robustness. Building upon the previously identified optimization difficulties induced by corrupted data, Table 1 and Fig. 7 further quantify the severe consequences of this training bottleneck at the inference level. The overall performance evaluation indicates that when test data is subjected to H.264 bitstream corruption, all State-of-the-Art (SOTA) models suffer from catastrophic accuracy degradation. As presented in Table 1 , the Relative Performance Drop (RPD) predominantly falls within the high range of 29% to 36%. Specifically, Faster R-CNN exhibits an RPD as high as 36.40%, while even the relatively robust RTMDet incurs a performance loss of 29.71%. Figure 7 (a) and (b) intuitively visualize this phenomenon via histograms, where a significant disparity exists between the blue bars (representing the clean source domain) and the orange bars (representing the corrupted target domain). This confirms that models trained exclusively on clean data are incapable of effectively resisting feature space distortions inherent in video transmission, demonstrating extremely weak generalization capabilities in corrupted environments. Table 1 Detection Performance and Robustness Evaluation of Algorithms on the ResQ-UAV Benchmark Algorithm GT Test Corrupted Test RPD AP AP50 APs AP AP50 APs AP AP50 APs RTMDet 0.525 0.796 0.406 0.369 0.622 0.273 29.71% 21.86% 32.76% Faster R-CNN 0.511 0.747 0.417 0.325 0.521 0.249 36.40% 30.25% 40.29% Cascade R-CNN 0.539 0.799 0.439 0.362 0.586 0.272 32.84% 26.66% 38.04% DINO 0.612 0.852 0.526 0.402 0.638 0.328 34.31% 25.12% 37.64% A more critical finding emerges from an in-depth analysis of scale-specific metrics. By comparing the APs data in Table 1 , it is evident that the Relative Performance Drop (RPD) values for small object detection across all models (32.76%–40.29%) are significantly higher than their respective declines in overall AP. This trend is visually corroborated in Fig. 7 (c), which illustrates that the "performance gap" for small objects is the widest among all algorithms. The physical mechanism underlying this phenomenon lies in the fact that the VisDrone dataset contains a vast number of tiny targets captured from long distances, which occupy an extremely low pixel proportion. Conversely, macroblock loss and blocking artifacts induced by H.264 bitstream errors typically possess fixed spatial dimensions. When the scale of these artifacts is comparable to or even larger than that of the targets, the corrupted regions directly occlude or completely destroy the edge and texture information of small objects, causing the detector to entirely lose its feature response capability for such targets. In conclusion, the experimental results demonstrate that deploying models trained via traditional clean-data paradigms directly into real-world UAV links subject to communication interference poses significant risks. Existing datasets and training strategies fail to encompass the complex signal distortions induced by bitstream corruption. This limitation not only validates the universality of the cross-domain performance degradation observed in this study but also underscores the urgency and engineering value of constructing the ResQ-UAV dataset and conducting targeted robustness research. Effectiveness Analysis of the ResQ-UAV Dataset To validate the effectiveness of the ResQ-UAV dataset in enhancing model robustness against interference, we substituted the original training set with the ResQ-UAV corrupted training set. The models were then retrained under identical hyperparameter settings and subsequently evaluated on the corrupted validation set. Table 2 presents a detailed direct comparison of model performance pre- and post-retraining. Experimental results indicate that by incorporating the ResQ-UAV dataset for fine-tuning, all evaluated models achieved significant performance recovery. Specifically, employing the Performance Gain (GAP) as a quantitative metric, the results demonstrate that GAP values for all models are positive. Notably, the AP50 for RTMDet and Faster R-CNN improved by 4.30% and 5.90%, respectively. This trend is visually corroborated in Fig. 8 , where the green bars representing retraining performance surpass the corrupted baseline (yellow bars) across the vast majority of metrics. It is worth noting that the ResQ-UAV dataset not only restored overall detection accuracy but, more critically, significantly improved perception capabilities for Small Objects. For instance, Faster R-CNN achieved a gain of 1.80% in the APs metric, indicating that the dataset effectively reconstructed fine-grained features that were previously obscured by compression noise. Table 2 Evaluation of Performance Recovery and Gain (GAP) Following Retraining with the ResQ-UAV Dataset Algorithm Corrupted Test Retrain Test GAP AP AP50 APs AP AP50 APs AP AP50 APs RTMDet 0.369 0.622 0.273 0.392 0.665 0.282 + 2.30% + 4.30% + 0.90% Faster R-CNN 0.325 0.521 0.249 0.336 0.580 0.267 + 1.10% + 5.90% + 1.80% Cascade R-CNN 0.362 0.586 0.272 0.380 0.616 0.286 + 1.80% + 3.00% + 1.40% DINO 0.402 0.638 0.328 0.413 0.673 0.314 + 1.10% + 3.50% -1.40% To further validate the quality of the ResQ-UAV dataset and its stability during the optimization process, we analyzed the convergence behavior of the models during retraining. Experimental observations revealed that when training on corrupted data, both the Classification Loss and Regression Loss exhibited smooth and rapid descent trends, with a complete absence of gradient explosion or oscillation throughout the entire process. This provides compelling evidence that while the generated ResQ-UAV dataset simulates authentic bitstream corruption, it preserves adequate semantic structural information to support gradient descent and feature learning within the models. With the increase in training epochs, the validation mAP demonstrated a steady ascent and ultimately converged near the model's theoretical upper bound on the clean dataset. This further corroborates that employing ResQ-UAV for targeted domain adaptation training constitutes an effective approach to resolving detection failures induced by UAV video transmission corruption. Usage Notes Data Accessibility and Standardization. The ResQ-UAV dataset is now fully accessible to the academic community via a public repository, intended to accelerate research and development in robust UAV vision systems. To maximize data usability and interoperability, we have implemented strict standardization regarding the directory structure and annotation formats. The dataset adopts a hierarchical storage architecture wherein high-quality Source Images and their corresponding Bitstream-Corrupted Images share identical file naming conventions and sequence identifiers. This ensures precise indexing and seamless transition between data from different domains. Furthermore, all object annotations adhere to the standard MS COCO JSON format. This implies that researchers can directly integrate this dataset into mainstream deep learning frameworks, such as MMDetection and Detectron2, for experimentation without the need for cumbersome format conversions. Potential Application Scenarios and Extended Value. Beyond serving as a benchmark for robust object detection, ResQ-UAV, leveraging its unique "pixel-level alignment" characteristic, exhibits significant potential for exploration across broader computer vision domains: Video Restoration and Enhancement : Given that the dataset provides pixel-level aligned "clean-corrupted" image pairs, it naturally constitutes an ideal training set for supervised learning tasks. Researchers can utilize pristine frames as pixel-level Ground Truth to train deep neural networks for tasks such as video De-blocking, Artifact Removal, and Super-Resolution Reconstruction, specifically targeting the complex non-linear distortions introduced by H.264 codec errors. Robust Multi-Object Tracking : Since the dataset fully preserves the temporal continuity of the original video sequences, it naturally extends the boundaries of research on temporal consistency. The availability of such sequential data enables the community to deeply analyze how Packet Loss and mosaic artifacts disrupt the continuity of motion trajectories. This opens new avenues for developing "resilient tracking algorithms"—algorithms capable of maintaining target identity (ID) stability even when visual features are intermittently lost due to transmission failures. Small Object Detection in Noisy Environments : Considering the prevalence of tiny objects within VisDrone scenarios, ResQ-UAV constitutes a highly challenging testbed. It not only challenges the model's feature extraction capabilities regarding small objects but also tests the limits of attention mechanisms in distinguishing foreground targets from background noise under extremely low Signal-to-Noise Ratio (SNR) conditions. This fosters the further evolution of Feature Pyramid Network (FPN) designs and context modeling techniques. Declarations Code availability To ensure research reproducibility and foster collaborative innovation in robust UAV vision systems, the ResQ-UAV dataset and its accompanying algorithmic benchmark code have been made fully open-source to the academic community, accessible at: https://github.com/dj-dengjian/ResQ-UAV.git. This repository hosts rigorously standardized data organization scripts and MS COCO-formatted annotation files, ensuring precise alignment in terms of filenames and sequence identifiers between pristine source images and bitstream-corrupted images. By providing these core resources, we aim not only to support benchmark evaluations for robust object detection tasks but also to establish an ideal experimental platform for video de-blocking, artifact removal, super-resolution reconstruction, and resilient multi-object tracking algorithms designed to handle transmission faults. Consequently, this initiative facilitates the exploration of boundaries in computer vision technology regarding complex noisy environments and small object perception. Acknowledgments This study was supported by the National Social Science Fund of China (No. 22&ZD169) and the Key project of Civil Aviation Joint Fund of National Natural Science Foundation of China (No. U2133207). Author contributions Author Jian Deng designed the experiments, generated the ResQ-UAV dataset, performed the benchmark evaluations, and wrote the original draft. Author Zeyu Liu assisted in data annotation and the implementation of the detection algorithms. Author Zihan Yu validated the experimental results. Author Honghai Zhang supervised the project, provided guidance and funding support, and reviewed and edited the paper. All authors reviewed the manuscript. competing interests The authors declare no competing interests. References Mester, B., Frieler, K., Korup, O., Desai, B. & Schewe, J. Socioeconomic Predictors of Vulnerability to Flood-Induced Displacement. Nat. Commun. 16, (2025). Pesonen, J. et al. Boreal Forest Fire: Uav-Collected Wildfire Detection and Smoke Segmentation Dataset. Sci. Data . 12, 1419 (2025). Deng, J., Zhang, H., Zhang, Y., Hua, M. & Sun, Y. A Method for Uav Path Planning Based On G-Maponet Reinforcement Learning. Drones. 9, 871 (2025). Deng, J., Zhang, H., Zhang, Y. & Sun, Y. Research On Trajectory Planning for a Limited Number of Logistics Drones (≤ 3) Based On Double-Layer Fusion Gwop. Drones. 9, 671 (2025). K. R., A. et al. Manipal-Uav Person Detection Dataset: A Step Towards Benchmarking Dataset and Algorithms for Small Object Detection. Isprs-J. Photogramm. Remote Sens. 195, 77–89 (2023). Simões, D. P., Oliveira, H. C. D. & Pereira, D. R. Unicamp-Uav: An Open Dataset for Human Detection in Uav Imagery. Isprs-J. Photogramm. Remote Sens. 231, 119–136 (2026). Robicquet, A., Sadeghian, A., Alahi, A. & Savarese, S. Learning Social Etiquette: Human Trajectory Understanding in Crowded Scenes. 14th European Conference on Computer Vision (ECCV 2016). Amsterdam, The Netherlands, 2016:549–565. A. V., S., Sankaran, P. & C. V., R. Fine-Tuned Deep Models for Niche Datasets — People Detection in Uav Building Images to Aid Rescue Operations. Int. J. Appl. Earth Obs. Geoinf. 145, 104985 (2025). Wiegand, T., Sullivan, G. J., Bjontegaard, G. & Luthra, A. Overview of the H.264/Avc Video Coding Standard. Ieee Trans. Circuits Syst. Video Technol. 13, 560–576 (2003). Ren, S., He, K., Girshick, R. & Sun, J. Faster R-Cnn: Towards Real-Time Object Detection with Region Proposal Networks. In: Cortes C, Lawrence ND, Lee DD, Sugiyama M, Garnett R, eds. ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 28 (NIPS 2015) . 29th Annual Conference on Neural Information Processing Systems (NIPS), 2015. Cai, Z. W., Vasconcelos, N. & IEEE. Cascade R-Cnn: Delving Into High Quality Object Detection. 2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR) . 31st IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018:6154–6162. Lyu, C. et al. Rtmdet: An Empirical Study of Designing Real-Time Object Detectors. Arxiv . abs/2212.07784, (2022). Zhang, H. et al. Dino: Detr with Improved Denoising Anchor Boxes for End-to-End Object Detection. International Conference on Learning Representations (ICLR) , 2023. Wu, X., Li, W., Hong, D., Tao, R. & Du, Q. Deep Learning for Unmanned Aerial Vehicle-Based Object Detection and Tracking: A Survey. Ieee Geosci. Remote Sens. Mag. 10, 91–124 (2022). Zhu, P. et al. Detection and Tracking Meet Drones Challenge. Ieee Trans. Pattern Anal. Mach. Intell. 44, 7380–7399 (2022). Gui-Song Xia, X. B. J. D. & Jiebo Luo, M. D. M. P. Dota: A Large-Scale Dataset for Object Detection in Aerial Images. 2018 Ieee/Cvf Conference On Computer Vision and Pattern Recognition . (2018). Yu, X., Gong, Y., Jiang, N., Ye, Q. & Han, Z. Scale Match for Tiny Person Detection. 2020 Ieee Winter Conference On Applications of Computer Vision (Wacv) . 1246–1254 (2019). Huszak, A. & Imre, S. Analysing Gop Structure and Packet Loss Effects On Error Propagation in Mpeg-4 Video Streams. 4th International Symposium on Communications, Control and Signal Processing : IEEE, 2010:1–5. Liu, T. et al. Bitstream-Corrupted Video Recovery: A Novel Benchmark Dataset and Method. Advances in Neural Information Processing Systems (NeurIPS) , 2023. Ghasempour, M. & Ghanbari, M. A Low Complexity System for Multiple Data Embedding Into H.264 Coded Video Bit-Stream. Ieee Trans. Circuits Syst. Video Technol. 30, 4009–4019 (2020). J., L. et al. A Comprehensive Benchmark for Single Image Compression Artifact Reduction. Ieee Trans. Image Process. 29, 7845–7860 (2020). Wiegand, T., Schwarz, H., Joch, A., Kossentini, F. & Sullivan, G. J. Rate-Constrained Coder Control and Comparison of Video Coding Standards. Ieee Trans. Circuits Syst. Video Technol. 13, 688–703 (2003). Stockhammer, T., Hannuksela, M. M. & Wiegand, T. H.264/Avc in Wireless Environments. Ieee Trans. Circuits Syst. Video Technol. 13, 657–673 (2003). Khawaja, W., Guvenc, I., Matolak, D. W., Fiebig, U. & Schneckenburger, N. A Survey of Air-to-Ground Propagation Channel Modeling for Unmanned Aerial Vehicles. Ieee Commun. Surv. Tutor. 21, 2361–2391 (2019). Kurutepe, E., Civanlar, M. R. & Tekalp, A. M. Client-Driven Selective Streaming of Multiview Video for Interactive 3Dtv. Ieee Trans. Circuits Syst. Video Technol. 17, 1558–1565 (2007). Seshadrinathan, K. & Bovik, A. C. Motion Tuned Spatio-Temporal Quality Assessment of Natural Videos. Ieee Trans. Image Process. 19, 335–350 (2010). Zhang, S. et al. Widerperson: A Diverse Dataset for Dense Pedestrian Detection in the Wild. Ieee Trans. Multimedia . 22, 380–393 (2020). Sakaridis, C., Dai, D. & Van Gool, L. Semantic Foggy Scene Understanding with Synthetic Data. Int. J. Comput. Vis. 126, 973–992 (2018). Teney, D., Liu, L., van den Hengel, A. & IEEE. Graph-Structured Representations for Visual Question Answering. 30TH IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2017) . 30th IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017:3233–3241. Lin, T. et al. Microsoft Coco: Common Objects in Context. In: Fleet D, Pajdla T, Schiele B, Tuytelaars T, eds. COMPUTER VISION - ECCV 2014, PT V . 13th European Conference on Computer Vision (ECCV), 2014:740–755. Kai, C. et al. Mmdetection: Open Mmlab Detection Toolbox and Benchmark [Arxiv]. Arxiv . 13 (2019). Campillos-Llanos, L. et al. Transformer-Based Relation Extraction and Concept Normalization Using an Annotated Clinical Trials Corpus. Sci. Data . (2026). Qian, N. On the Momentum Term in Gradient Descent Learning Algorithms. Neural Netw. 12, 145–151 (1999). Sun, S., Cao, Z., Zhu, H. & Zhao, J. A Survey of Optimization Methods From a Machine Learning Perspective. Ieee T. Cybern. 50, 3668–3681 (2020). Yang, L. & Shami, A. On Hyperparameter Optimization of Machine Learning Algorithms: Theory and Practice. Neurocomputing . 415, 295–316 (2020). Tan, M., Pang, R. & Le, Q. V. Efficientdet: Scalable and Efficient Object Detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2020. Liu, J. et al. Benchmarking Object Detection Robustness Against Real-World Corruptions. Int. J. Comput. Vis. 132, 4398–4416 (2024). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8741863","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":594840709,"identity":"a69771eb-0df7-4f92-bced-a9ccbb7655dc","order_by":0,"name":"Jian Deng","email":"","orcid":"","institution":"Nanjing University of Aeronautics and Astronautics","correspondingAuthor":false,"prefix":"","firstName":"Jian","middleName":"","lastName":"Deng","suffix":""},{"id":594840710,"identity":"4328178e-f992-40cb-abd1-2469781a5741","order_by":1,"name":"Honghai Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABFUlEQVRIiWNgGAWjYBACAyBmZmCQACMGhgqoMA/xWs4Qr4UBooWxjQgt5uy9h18X1FgwyM9ufvbw67w7if3tBxgfvG1jkDfHocWy51ya9YxjEgyMc46ZG8tue5Y440wCs+HcNgbDnQ04HHYjx8yYh02CgVkiwUxactvhxIYDCWzSvG0MCQYH8Gn5J8HAJpH+TVpyzuHE+ecfsP8moMX4MW+bBAOPRI6Z5MeGw4kbbiSwMePVcuaMGTNvnwSDhEROmTTDscPGG288bJacc07CcAMuLcd7jD/zfKtjkJ+Rvk3yR81h2Xnnkw9+eFNmI4/LFiBgA8VIfQOQYIZEByOILYFTPUjhBxiL8Qc+daNgFIyCUTBiAQDvnVmfvszvwwAAAABJRU5ErkJggg==","orcid":"","institution":"Nanjing University of Aeronautics and Astronautics","correspondingAuthor":true,"prefix":"","firstName":"Honghai","middleName":"","lastName":"Zhang","suffix":""},{"id":594840711,"identity":"3369067d-6e08-473d-bb2c-2e8bdf5d58d5","order_by":2,"name":"Zeyu Liu","email":"","orcid":"","institution":"Nanjing University of Aeronautics and Astronautics","correspondingAuthor":false,"prefix":"","firstName":"Zeyu","middleName":"","lastName":"Liu","suffix":""},{"id":594840712,"identity":"4aabeda4-2704-42f1-b5f4-c6084e53a927","order_by":3,"name":"Zihan Yu","email":"","orcid":"","institution":"Nanjing University of Aeronautics and Astronautics","correspondingAuthor":false,"prefix":"","firstName":"Zihan","middleName":"","lastName":"Yu","suffix":""}],"badges":[],"createdAt":"2026-01-30 13:53:57","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8741863/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8741863/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":103504723,"identity":"03f94d47-093c-4906-856c-373802ae452d","added_by":"auto","created_at":"2026-02-26 13:21:07","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":941193,"visible":true,"origin":"","legend":"\u003cp\u003eVisualization of corrupted imagery from the ResQ-UAV dataset dedicated to future rescue research. The figure encompasses bitstream corruption degradation characteristics across diverse scene perspectives, aiming to provide a foundational data resource for the development of robust recognition technologies tailored to complex disaster environments.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/52d5cea21e377d339ce3e5bd.png"},{"id":103199403,"identity":"1106c8f4-968a-47a8-8bfa-ad97fa87f011","added_by":"auto","created_at":"2026-02-23 05:35:07","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":44843,"visible":true,"origin":"","legend":"\u003cp\u003eData Storage Structure.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/c0eda125e78352a72e9e5ed0.png"},{"id":103505816,"identity":"b08c4591-8b16-4007-9552-f5dc97bd9f42","added_by":"auto","created_at":"2026-02-26 13:33:08","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":186088,"visible":true,"origin":"","legend":"\u003cp\u003eVisual analysis of the training process and statistical characteristics of the RTMDet model in the clean source domain versus the corrupted target domain: (a) Trajectories of training accuracy improvement, (b) Statistical distribution of mAP metrics, (c) Convergence curves of training loss, and (d) Distributional characteristics of training loss values.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/24da907df17a01cd13342476.png"},{"id":103199404,"identity":"d5dafe85-a3fe-4929-b536-695eba4939db","added_by":"auto","created_at":"2026-02-23 05:35:07","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":158509,"visible":true,"origin":"","legend":"\u003cp\u003eVisual analysis of the training process and statistical characteristics of the Faster R-CNN model in the clean source domain versus the corrupted target domain: (a) Trajectories of training accuracy improvement, (b) Statistical distribution of mAP metrics, (c) Convergence curves of training loss, and (d) Distributional characteristics of training loss values.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/ee71056160ee0f7c593f6650.png"},{"id":103199410,"identity":"a3eb89e1-9791-4a30-966b-5d15640f6a44","added_by":"auto","created_at":"2026-02-23 05:35:07","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":179894,"visible":true,"origin":"","legend":"\u003cp\u003eVisual analysis of the training process and statistical characteristics of the Cascade R-CNN model in the clean source domain versus the corrupted target domain: (a) Trajectories of training accuracy improvement, (b) Statistical distribution of mAP metrics, (c) Convergence curves of training loss, and (d) Distributional characteristics of training loss values.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/ed914c7104a386c51541a412.png"},{"id":103199407,"identity":"2b096bc2-695b-4561-87a5-19635cbe5e46","added_by":"auto","created_at":"2026-02-23 05:35:07","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":160689,"visible":true,"origin":"","legend":"\u003cp\u003eVisual analysis of the training process and statistical characteristics of the DINO model in the clean source domain versus the corrupted target domain: (a) Trajectories of training accuracy improvement, (b) Statistical distribution of mAP metrics, (c) Convergence curves of training loss, and (d) Distributional characteristics of training loss values.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/e3fe79655808e3c282dcd818.png"},{"id":103505084,"identity":"d3094947-631b-4943-82fc-4e5ead28940f","added_by":"auto","created_at":"2026-02-26 13:23:25","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":92528,"visible":true,"origin":"","legend":"\u003cp\u003eVisual comparison of cross-domain performance discrepancies for four mainstream detection models between the clean benchmark (GT Test) and the corrupted test set (Corrupted Test): (a) Cross-domain comparison of Average Precision (AP), (b) Comparison of Average Precision under a loose threshold (AP50), and (c) Cross-domain comparison of small object detection precision (APs).\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/cc5822d1485d1ab7c7aad077.png"},{"id":103505876,"identity":"8faf5378-2d13-478c-b0a1-c2984bf5ad1c","added_by":"auto","created_at":"2026-02-26 13:33:20","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":89388,"visible":true,"origin":"","legend":"\u003cp\u003eVisual comparison of performance recovery and enhancement effects following retraining with the ResQ-UAV dataset: (a) Comparison of Average Precision (AP) recovery, (b) Comparison of Average Precision improvement under a loose threshold (AP50), and (c) Comparison of restoration and improvement in small object detection precision (APs).\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/230f0a92ccc73c4e34a220c1.png"},{"id":107707625,"identity":"636f09a7-6c5c-4653-bece-3b3f00955f24","added_by":"auto","created_at":"2026-04-24 09:20:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2185201,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8741863/v1/0bc235a3-80b5-461d-80c1-b6909d67e778.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"\u003cp\u003eResQ-UAV: A Novel Dataset Supporting Robust Recognition for Future Drone Rescue Missions\u003c/p\u003e","fulltext":[{"header":"Background \u0026 Summary","content":"\u003cp\u003eIn recent years, the frequent occurrence of natural disasters globally\u0026mdash;such as floods, earthquakes, and typhoons\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e\u0026mdash;has posed significant threats to human life and property. During the critical \"golden rescue window\" following such catastrophes, the complex and dynamic post-disaster environment\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e often leads to the paralysis of ground transportation, communication interruptions, and risks of secondary disasters, thereby imposing severe impediments to traditional manual search and rescue operations.\u003c/p\u003e \u003cp\u003eAgainst this backdrop, Unmanned Aerial Vehicles (UAVs)\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e, leveraging their high mobility, flexible deployment, and extensive aerial field of view, have demonstrated broad application prospects in post-disaster emergency response and humanitarian aid. Particularly in Search and Rescue (SAR) missions, UAVs\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e can rapidly penetrate areas that are inaccessible or hazardous to humans. They facilitate high-efficiency, all-weather scanning of large-scale regions, thereby enabling the rapid localization of trapped individuals and significantly enhancing rescue response speed and success rates.\u003c/p\u003e \u003cp\u003eTo address the specific challenges of UAV visual perception, the academic community has recently introduced several targeted, high-quality datasets. Addressing the challenge of \"small objects and high density,\" Manipal-UAV (2023)\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e provides large-scale annotated data containing over 150,000 instances, specifically designed to evaluate model detection capabilities under extreme scale variations. Furthermore, Unicamp-UAV (2026)\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e incorporates a negative sample mechanism to focus on the issue of high false-positive rates in open environments, representing the latest standards in data acquisition and annotation. Regarding \"specific rescue scenarios,\" the Stanford Drone Dataset (2020)\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e offers trajectory analysis data from a top-down aerial perspective, while the recent NITC (2025)\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e focuses on person detection against building backgrounds, aiming to resolve recognition difficulties caused by occlusion from structural elements.\u003c/p\u003e \u003cp\u003eAlthough existing datasets have enriched UAV vision research, the field still faces severe challenges when applied to realistic post-disaster rescue scenarios, primarily manifested in the following aspects:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eScarcity of data with video transmission link impairments\u003c/b\u003e: Existing UAV datasets typically focus on environmental challenges at the optical acquisition end, such as fog, rain, or low-light conditions. However, in actual Beyond Visual Line of Sight (BVLOS) rescue or logistics missions, limited by bandwidth fluctuations and electromagnetic interference, the image transmission link often suffers from specific bitstream corruptions (e.g., mosaic effects and frame loss caused by H.264/H.265 decoding errors\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e). Most currently public datasets consist of locally stored lossless images, lacking corrupted samples that reflect real-world communication link failures. Consequently, models trained on existing data are prone to generalization failure when encountering real-time transmission artifacts, leading to severe missed detections.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDiscrepancy between the ideal training domain and the real-world deployment domain\u003c/b\u003e: The performance of deep learning models relies heavily on the consistency between training and testing data distributions. Existing benchmark models are typically trained on clear, high-quality \"source domain\" data, assuming that test scenarios contain only simple linear noise (e.g., Gaussian blur). However, real-world video transmission impairments exhibit highly non-linear and structured characteristics (such as blocking artifacts), constituting a significant \"domain shift.\" Due to the lack of adversarial training data specifically targeting transmission-corrupted scenarios, existing research struggles to effectively evaluate and improve model robustness under this specific domain discrepancy.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eTo bridge the gap between optical imaging and video transmission stability assessment, we constructed and open-sourced the ResQ-UAV dataset, designed for future rescue research (as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). Unlike traditional acquisition methods, the core construction strategy of ResQ-UAV involves processing existing high-quality UAV data by simulating authentic H.264 video coding transmission errors. This process generates corrupted samples possessing typical degradation characteristics associated with bitstream corruption.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the ResQ-UAV dataset provides a visual representation of bitstream corruption degradation characteristics across diverse scene perspectives. These images faithfully reproduce the visual impairments commonly encountered in communication-constrained disaster sites, including non-linear noise such as severe blocking artifacts, color distortion, local pixel loss, and image tearing. The visualization results in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e clearly demonstrate how such structured degradation blurs the edge features of person targets, leading to indistinguishability between targets and the background. Consequently, this provides a challenging and indispensable foundational data resource for developing robust recognition technologies tailored to complex disaster environments.\u003c/p\u003e \u003cp\u003eThe ResQ-UAV dataset comprises 16,414 frames derived from 32 video sequences, spanning a spectrum of transmission states from lossless to severely corrupted. The dataset contains a total of 282,095 meticulously cleaned bounding boxes for persons. The scenarios encompass a variety of complex environments, including urban arterials, intersections, suburban areas, and playgrounds, covering both diurnal and nocturnal periods. To facilitate research on domain adaptation algorithms, we constructed a paired \"clean-corrupted\" data structure and performed semantic merging and unified annotation for the \"pedestrian\" and \"person\" categories within the original data.\u003c/p\u003e \u003cp\u003eTo assess the challenging nature of the ResQ-UAV dataset in the field of robustness detection, we trained and evaluated four representative state-of-the-art (SOTA) object detection algorithms: the classic Faster R-CNN\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e, the multi-stage Cascade R-CNN\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, the latest real-time detector RTMDet\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e, and the Transformer-based DINO\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Experimental results indicate that, compared to their performance on clean datasets, existing SOTA methods exhibit a marked degradation in performance (a substantial attenuation in mAP) on ResQ-UAV. This decline is primarily attributed to the prevalent feature loss caused by bitstream corruption within ResQ-UAV, demonstrating that current detectors lack the feature extraction capabilities and interference resilience necessary to handle such structured noise.\u003c/p\u003e \u003cp\u003eAs an exploratory initiative oriented towards disaster rescue scenarios, ResQ-UAV establishes a UAV person detection benchmark focused on evaluating video transmission robustness. By simulating bitstream corruption characteristics inherent to real-world communication-constrained environments, this dataset augments the sample diversity of existing UAV vision research under complex transmission conditions. It not only provides essential test data for research on small person detection in low-bitrate and high-packet-loss environments but also serves as a critical data foundation and validation platform for developing UAV Search and Rescue (SAR)\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e vision systems capable of adapting to complex channel interference with enhanced transmission robustness.\u003c/p\u003e"},{"header":"Materials","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eDate sources\u003c/h2\u003e \u003cp\u003eTo construct a dataset of corrupted videos with high realism and significant challenge, this study selected the VisDrone2019-VID\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e dataset as the raw video benchmark. VisDrone stands as one of the largest and most authoritative benchmarks in the field of UAV visual analysis. The rationale for selecting it as the foundation for ResQ-UAV is primarily based on the following three considerations:\u003c/p\u003e \u003cp\u003e \u003cb\u003eFirst, the temporal continuity of video sequences.\u003c/b\u003e Unlike datasets based on static images such as DOTA\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e or TinyPerson\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, VisDrone-VID provides complete video sequences. This characteristic is critical for simulating H.264 bitstream corruption. Since video coding relies on inter-frame prediction and Group of Pictures (GOP) structures\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e, single static images cannot capture the temporal propagation effects of decoding errors. The continuous frame structure of VisDrone enables us to authentically reproduce the non-linear dynamic degradation process where \"corruption in a preceding frame leads to artifacts in subsequent frames.\"\u003c/p\u003e \u003cp\u003e \u003cb\u003eSecond, extremely rich scene diversity.\u003c/b\u003e The original dataset was collected from 14 different cities across China\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e, spanning thousands of kilometers geographically. It covers a wide range of environments, including urban arterials, busy intersections, quiet suburbs, playgrounds, and residential areas. This extensive coverage ensures the complexity of background textures, effectively evaluating model robustness when background clutter is superimposed with corruption artifacts, thereby preventing the model from overfitting to a single background type.\u003c/p\u003e \u003cp\u003e \u003cb\u003eFinally, authentic UAV perspective challenges.\u003c/b\u003e The data acquisition encompasses various lighting conditions (daytime, nighttime, overcast) as well as different shooting angles (top-down, oblique) and flight altitudes. This implies that the original videos inherently contain UAV-specific challenges such as motion blur and drastic scale variations. Superimposing bitstream corruption on top of these inherent optical challenges allows for the construction of a \"double difficulty\" test benchmark that most closely approximates real-world post-disaster search and rescue scenarios.\u003c/p\u003e \u003cp\u003eIn summary, the video stream characteristics and high scene complexity of VisDrone-VID make it the ideal substrate for generating bitstream corruption benchmarks. Consequently, this study selected 32 video clips containing abundant human activities from the dataset as the core data source.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eData processing\u003c/h3\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eSelection of Corruption Scheme\u003c/h2\u003e \u003cp\u003eIn the construction of a corrupted video dataset, the selection of the data generation scheme is critical, as it directly determines the realism and challenge of the evaluation benchmark. Existing research on video inpainting and error concealment\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e predominantly adopts traditional \"mask simulation\" strategies, which involve directly overlaying error masks of predefined shapes onto decoded video frames. These masks typically manifest as regular slices or rectangular blocks, with positions and dimensions that are often fixed or follow simple random distributions.\u003c/p\u003e \u003cp\u003eHowever, such pixel-level simulation methods possess notable limitations, failing to capture the physical characteristics of video data loss inherent in real-world multimedia communications. In actual UAV image transmission links, video streams undergo compression via standards such as H.264/AVC. Consequently, packet loss or signal interference during transmission occurs within the underlying \"bitstream domain\" rather than the \"pixel domain\"\u003csup\u003e20\u003c/sup\u003e. Governed by the inherent inter-frame prediction mechanisms and Context-Adaptive Variable Length Coding (VLC/CABAC) characteristics of H.264 encoding, minor perturbations in the bitstream layer induce highly complex visual artifacts at the decoder, exhibiting non-linear propagation characteristics. In contrast to simple artificial masks, errors precipitated by bitstream corruption are more diverse in typology. They encompass severe blocking artifacts, color artifacts\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e, texture loss, and global or local frame misalignment resulting from damaged motion vectors. Therefore, to bridge the gap between laboratory simulations and real-world operational failures, this study adopts a bitstream-based corruption scheme to generate corrupted samples that possess both high realism and stochasticity.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eBitstream Corruption Model\u003c/h3\u003e\n\u003cp\u003eTo generate corrupted data that is both controllable and possesses high realism, this study develops a bitstream corruption methodology based on the underlying syntax of video coding. Distinct from traditional pixel-level masking, this method adheres to the error control mechanisms of the H.264/AVC standard within IP network transmission\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e, constructing a universal framework capable of precisely simulating real-world network packet loss and bit errors. Specifically, leveraging the Network Abstraction Layer (NAL) structure and error propagation characteristics of H.264\u003csup\u003e23\u003c/sup\u003e, we propose a three-parameter stochastic corruption model (P, L, S), utilizing this tuple to precisely define the patterns and intensity of the degradation:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eFrame Corruption Probability (P)\u003c/b\u003e: This parameter simulates the packet loss rate in unreliable channels and dictates the frequency of corrupted frames within the video sequence, representing the statistical proportion of target frames randomly selected for corruption within a Group of Pictures (GOP).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eCorruption Start Position (L)\u003c/b\u003e: Refers to the byte offset within the payload of the selected Network Abstraction Layer Unit (NALU) where error injection or data truncation occurs.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eCorrupted Segment Length (S)\u003c/b\u003e: Defines the length of the continuous bitstream data (in bytes) to be removed or tampered with, used to simulate varying degrees of burst errors.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe implementation logic of this method is strictly grounded in the parsing of the H.264/AVC bitstream syntax structure\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. A standard H.264 bitstream is composed of a series of NALUs, wherein key control information\u0026mdash;such as the Sequence Parameter Set (SPS), Picture Parameter Set (PPS), and Slice Header\u0026mdash;constitutes a negligible proportion (\u0026lt;\u0026thinsp;0.01%) of the total bitstream but is vital for decoding synchronization. To prevent decoder crashes or catastrophic sequence disruptions, the model adopts an Unequal Error Protection (UEP) strategy. This involves bypassing high-sensitivity header information and targeting only the dominant Video Coding Layer (VCL) data for specific segment removal operations. Through this low-level manipulation, the generated corrupted video not only retains a valid syntax structure but also faithfully replicates a variety of highly non-linear, realistic visual impairments, including blocking artifacts, texture loss, and motion vector misalignment.\u003c/p\u003e\n\u003ch3\u003eParameter Configuration and Adaptive Adjustment\u003c/h3\u003e\n\u003cp\u003eAlthough this study adopts a general bitstream corruption framework, given the substantial disparity in spatial scale between ResQ-UAV (derived from VisDrone) and traditional low-resolution datasets, we have performed a targeted adaptive reconfiguration of the corruption parameters by integrating video transmission characteristics with visual perception theory:\u003c/p\u003e \u003cp\u003e \u003cb\u003eCorruption Probability (P): Realistic Fitting of Communication Links.\u003c/b\u003e In this study, the frame corruption probability is set to P\u0026thinsp;=\u0026thinsp;1/16, implying that, on average, one frame within a Group of Pictures (GOP) of length 16 undergoes bitstream corruption. This threshold setting references the realistic fading and packet loss models of UAV Air-to-Ground (A2G) propagation channels \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. This ratio constitutes a critical equilibrium point: on one hand, it is sufficient to simulate significant visual interference generated under unreliable channel conditions; on the other hand, it adheres to the stability constraints of streaming media transmission, avoiding decoder resets or complete video stream interruptions caused by excessive error rates (Packet Loss Rate\u0026thinsp;\u0026gt;\u0026thinsp;10%). Consequently, this accurately reflects the characteristics of real-world scenarios where UAVs experience intermittent packet loss in weak network environments.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCorruption Position (L): Intensifying Cascaded Error Propagation.\u003c/b\u003e To maximize the attack efficacy of a single corruption event, we set the corruption start position to L\u0026thinsp;=\u0026thinsp;0.2, meaning that error injection is primarily concentrated within the first 20% of the NALU payload. According to quantitative analyses of the impact of packet loss on subjective video quality\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e, the initial segment of a NALU typically carries critical decoding syntax elements such as Motion Vector Differences (MVD) and Macroblock Modes (MbMode). Corruption at an early position not only directly destroys subsequent data within the current frame but also leverages the inter-frame prediction mechanism of H.264 to trigger a severe \"avalanche effect,\" causing temporal error diffusion to subsequent dependent frames (P-frames/B-frames). This strategic position selection significantly enhances the persistence and destructive power of artifacts, thereby substantially increasing the difficulty of robustness testing for detection algorithms.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCorruption Size (S): Resolution-Based Perceptual Scale Augmentation.\u003c/b\u003e The parameter augmentation targeting high-resolution video represents the most significant adjustment in this study. Existing research on corrupted video is predominantly based on low-resolution materials (480P or 720P), where the commonly used corruption size of 4096 bytes produces only negligible visual noise in the 4K (3840\u0026times;2160) Ultra-High-Definition (UHD) bitstreams widely adopted in the VisDrone dataset, failing to constitute effective interference. According to multi-scale perception theory in Spatio-Temporal Video Quality Assessment (ST-VQA)\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e, the saliency of visual artifacts is highly dependent on their proportion relative to the overall image. Therefore, to ensure that the corruption effects maintain strong visual impact and detection challenges within 4K frames, we substantially increased the corruption size to S\u0026thinsp;=\u0026thinsp;20,480 bytes. This magnitude adjustment ensures that the blocking artifacts and signal tearing in ResQ-UAV match the scale of high-resolution UAV imagery, thereby enabling effective evaluation of model generalization capabilities under extreme conditions characterized by \"large field of view, small objects, and strong interference.\"\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eObject annotation\u003c/h2\u003e \u003cp\u003eTo construct a benchmark suitable for high-precision personnel search and rescue as well as surveillance assessment, this study implemented a rigorous three-stage data processing pipeline consisting of \"Reconstruction\u0026ndash;Mapping\u0026ndash;Cleaning.\" This workflow aims to resolve the semantic ambiguity inherent in the original labels and ensure the spatial precision and visual validity of the annotated data within the Corrupted Domain.\u003c/p\u003e \u003cp\u003e \u003cb\u003eSemantic Reconstruction and Full Re-annotation of Source Domain Data\u003c/b\u003e. Although the original VisDrone dataset provides rich annotations, there exists significant overlap and ambiguity in the semantic definitions between the \"pedestrian\" and \"people\" categories. To eliminate the potential interference of label noise on subsequent model evaluation, this study discarded the original label system, referencing consistent annotation standards used in wild crowded pedestrian detection\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. Utilizing the intelligent annotation tool X-AnyLabeling, we conducted a comprehensive manual re-annotation from scratch on the 32 selected high-definition raw video sequences. In this process, all human targets were uniformly categorized into a single class, \"Person,\" thereby ensuring high semantic consistency and purity of the Ground Truth data in the source domain.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAutomatic Projection Mapping of Cross-Domain Labels.\u003c/b\u003e Upon establishing high-quality annotations for the source domain, we developed a semi-automated transfer workflow from the original domain to the corrupted domain. Adhering to the principle of geometric invariance in synthetic data domain adaptation research\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e\u0026mdash;which posits that H.264 bitstream corruption affects only image texture without altering the physical spatial coordinates of objects\u0026mdash;we directly projected and mapped the re-annotated bounding boxes from the source domain onto the corresponding ResQ-UAV corrupted video sequences. This strategy guarantees pixel-level alignment of spatial positions between Clean Samples and Corrupted Samples, providing a perfect comparative baseline for subsequently quantifying the performance degradation of algorithms under identical scenes but varying image quality conditions.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCleaning Verification Based on Visual Perceptibility.\u003c/b\u003e Mere coordinate mapping may retain targets that are visually completely lost, thereby violating the general criterion of \"Visual Perceptibility\" in dataset construction\u003csup\u003e\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. Therefore, following the completion of automatic mapping, we introduced a rigorous manual cleaning phase. Annotators used X-AnyLabeling to perform frame-by-frame verification on the corrupted videos, focusing on the removal of targets that became completely unrecognizable to the naked eye due to frame loss, extensive mosaic occlusion, or severe color distortion. Through this physical filtering mechanism, the final annotation set generated for ResQ-UAV retains precise spatial localization while effectively excluding visually imperceptible invalid samples, thus faithfully reflecting the effective boundaries of detection algorithms in extreme visual environments.\u003c/p\u003e \u003c/div\u003e"},{"header":"Data Records","content":"\u003cp\u003eThe ResQ-UAV dataset is hosted on a GitHub repository (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/dj-dengjian/ResQ-UAV.git\u003c/span\u003e\u003cspan address=\"https://github.com/dj-dengjian/ResQ-UAV.git\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), providing open access to the research community. Designed to maximize usability and support paired experimental analysis, the dataset features a highly structured hierarchical directory organization. The detailed file storage architecture and branching logic are depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe root directory of the dataset comprises three core components: the data folder for storing image sequences, the annotations folder for storing detection labels, and the mmdetection codebase integrated with benchmark detection algorithms. Detailed descriptions of each component are provided below:\u003c/p\u003e \u003cp\u003e \u003cb\u003eImage Data Organization (data).\u003c/b\u003e The data directory adopts a strict dual-branch parallel architecture to ensure a one-to-one correspondence between the clean source domain and the corrupted target domain:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003egt branch\u003c/b\u003e: Stores high-resolution lossless video frames extracted from the original VisDrone dataset, serving as the reference benchmark for visual perception.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ecorrupted branch\u003c/b\u003e: Stores corrupted video frames generated based on the H.264 bitstream corruption model, serving as the target data for robustness evaluation.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThese two branches share an identical subdirectory structure, both indexed by video sequence ID. Regarding data partitioning, the 32 video sequences are strictly divided into three mutually exclusive subsets: train (containing 22 sequences for model training), val (containing 4 sequences for hyperparameter validation), and test (containing 6 sequences for final performance evaluation). All images are stored in JPG format and are accompanied by frame-level metadata for traceability.\u003c/p\u003e \u003cp\u003e \u003cb\u003eAnnotation Data Organization (annotations).\u003c/b\u003e To accommodate the requirements of supervised learning and evaluation within corrupted environments, the annotations directory provides two sets of tailored ground truth labels, both organized in the standard MS COCO JSON format\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eannotations_gt\u003c/b\u003e: Contains high-quality ground truth labels for the \"Person\" category following the comprehensive re-annotation of pristine original images. These are utilized for training \"upper bound\" models or conducting source domain supervision.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eannotations_corrupted\u003c/b\u003e: Contains ground truth labels for corrupted images processed through the \"projection mapping\u0026ndash;manual cleaning\" pipeline. This annotation set excludes targets that are visually indiscernible within severely corrupted fields of view, thereby ensuring the objectivity of the evaluation.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eTo facilitate experimental configuration, annotation information for all subsets (Train/Val/Test) has been merged into independent JSON files (e.g., train_merged.json). Researchers can seamlessly transition between \"clean supervision\" and \"corrupted supervision\" experimental settings by simply modifying the file paths in the configuration file.\u003c/p\u003e"},{"header":"Technical Validation","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eModel Environment\u003c/h2\u003e \u003cp\u003eAll experiments were conducted on a high-performance computing workstation equipped with an 8-core processor, 32GB of RAM, and an NVIDIA Tesla V100 GPU with 16GB of video memory. The software environment was built upon Python 3.8, utilizing the open-source MMDetection framework\u003csup\u003e\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e for the entire pipeline of model training and inference.\u003c/p\u003e \u003cp\u003eTo comprehensively evaluate the robustness of detection architectures within corrupted video environments, this study selected four representative algorithms encompassing distinct paradigms: the classic two-stage detector Faster R-CNN based on ResNet-50; the multi-stage cascade architecture Cascade R-CNN, focused on enhancing detection quality at high IoU thresholds; the high-performance one-stage anchor-free detector RTMDet integrated with a CSPNeXt backbone; and the Transformer-based\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e end-to-end detection paradigm DINO.\u003c/p\u003e \u003cp\u003eRegarding training strategies, all models were initialized with official pre-trained weights from the MS COCO 2017 dataset to accelerate convergence and enhance the generalization capability of feature extraction. Tailored to the optimization characteristics of different architectures, the experiments employed either SGD\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e or AdamW\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e\u003c/sup\u003e optimizers, while uniformly adopting the Cosine Annealing algorithm\u003csup\u003e\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e for learning rate decay. This strategy, through periodic warm restarts and a smooth descent mechanism, effectively assists the model in circumventing local optima within complex loss surfaces, thereby facilitating convergence towards the global optimal solution.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eEvaluation Metrics\u003c/h2\u003e \u003cp\u003eTo comprehensively assess the holistic performance of models under UAV perspectives and bitstream-corrupted environments, this study constructs an integrated evaluation framework encompassing detection accuracy, training dynamics, and computational efficiency.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDetection Accuracy and Multi-Scale Perception Metrics.\u003c/b\u003e We strictly adhere to the evaluation protocols of the MS COCO object detection benchmark, adopting Average Precision (AP) as the core evaluation metric. Specifically, we report AP, defined as the mean Average Precision calculated over Intersection over Union (IoU) thresholds from 0.50 to 0.95 (with a step size of 0.05). This metric reflects the comprehensive performance of the model under varying degrees of localization rigor and serves as the primary criterion for measuring detector robustness. Furthermore, given the predominance of small objects in UAV datasets and their susceptibility to loss following bitstream corruption, we introduce fine-grained scale-specific evaluation metrics:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAP50\u003c/b\u003e: The average precision calculated at a fixed IoU threshold of 0.50. In the field of UAV vision, due to the extremely small size of targets and complex backgrounds, high IoU matching is often excessively stringent. Therefore, AP50} is widely regarded as a critical benchmark metric for assessing Small Object Detection performance.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAPs\u003c/b\u003e: The average precision specifically calculated for small objects (pixel area\u0026thinsp;\u0026lt;\u0026thinsp;32\u0026sup2;). In corrupted video detection tasks, the fluctuation of APs is particularly critical because blocking artifacts and ringing artifacts introduced by H.264 encoding often submerge the geometric features of tiny objects first, leading to detection failure.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eTraining Dynamics and Stability Analysis.\u003c/b\u003e To verify the model's fitting degree and generalization boundaries regarding corrupted data, we conduct an in-depth analysis of convergence behavior during training. By visualizing the evolution trends of Classification Loss, Regression Loss, and mAP across both training and validation sets throughout the iteration process, we can effectively monitor risks of overfitting or underfitting. Particularly when comparing Transformer architectures (e.g., DINO) with CNN architectures (e.g., Faster R-CNN), analyzing differences in their convergence speeds and final steady states helps reveal disparities in learning efficiency and robustness regarding different inductive biases when processing non-linear signal noise.\u003c/p\u003e \u003cp\u003e \u003cb\u003eComputational Complexity and Deployment Feasibility.\u003c/b\u003e Given that UAV onboard computing platforms are typically limited by strict Size, Weight, and Power (SWaP) constraints, the spatiotemporal complexity of the model is key to determining the feasibility of its engineering implementation. Consequently, we introduce Parameters and Floating Point Operations (FLOPs) to quantify the model's static spatial footprint and dynamic computational load. Simultaneously, to assess real-time processing capabilities, we test the single-frame inference Latency and Frames Per Second (FPS) of each model under a unified hardware environment. By constructing the Pareto Frontier\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e of accuracy (mAP) versus speed (FPS), we aim to identify the optimal equilibrium point that satisfies real-time response requirements while ensuring high-precision detection.\u003c/p\u003e \u003cp\u003eFinally, to quantify the specific impact of bitstream corruption on the detection system, beyond directly observing absolute accuracy on the corrupted dataset, this study introduces the Relative Performance Drop (RPD) as a robustness evaluation metric\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. Drawing from methods in image quality assessment regarding the resilience of deep networks against distortion, this metric intuitively reflects the model's resilience against interference during video transmission by calculating the relative magnitude of change in evaluation values between the original validation set (Clean) and the corrupted validation set (Corrupted). A lower RPD value indicates superior performance retention during image quality degradation, suggesting that the feature extraction network possesses stronger robustness against non-natural artifacts. This is of significant importance for evaluating algorithm reliability in extreme environments such as post-disaster search and rescue. The RPD formula is defined as follows:\u003cdiv id=\"Equ1\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ1\" name=\"EquationSource\"\u003e\n$${\\text{RPD}}=({\\text{A}}{{\\text{P}}_{{\\text{gt}}}} - {\\text{A}}{{\\text{P}}_{{\\text{corrupted}}}})/{\\text{A}}{{\\text{P}}_{{\\text{gt}}}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e1\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTo evaluate the performance recovery of the model following retraining with the introduced ResQ-UAV dataset, we propose the Performance Gain (GAP) metric. GAP is defined as the difference between the precision achieved by the model after training on corrupted data (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{A}}{{\\text{P}}_{{\\text{retrained}}}}\\)\u003c/span\u003e\u003c/span\u003e) and the baseline precision obtained when trained exclusively on clean data (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\({\\text{A}}{{\\text{P}}_{{\\text{corrupted}}}}\\)\u003c/span\u003e\u003c/span\u003e). A positive GAP value directly quantifies the effectiveness of the proposed dataset in mitigating Domain Shift and enhancing the detector's adaptability to corrupted signals, serving as a pivotal criterion for validating the core contribution of this study. The formula for GAP is defined as follows:\u003cdiv id=\"Equ2\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equ2\" name=\"EquationSource\"\u003e\n$${\\text{GAP}}={\\text{A}}{{\\text{P}}_{{\\text{retrained}}}} - {\\text{A}}{{\\text{P}}_{{\\text{corrupted}}}}$$\u003c/div\u003e\u003cdiv class=\"EquationNumber\"\u003e2\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eAnalysis of Results\u003c/h2\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003eRobustness Analysis of Existing Models\u003c/h2\u003e \u003cp\u003eTo assess the generalization capability of mainstream object detection algorithms within video transmission-corrupted scenarios, we established an evaluation benchmark based on \"source-domain supervision.\" In this experiment, four representative models\u0026mdash;RTMDet, Faster R-CNN, Cascade R-CNN, and DINO\u0026mdash;underwent parameter optimization utilizing exclusively the original VisDrone training set. Subsequently, inference evaluations were executed on both the clean original test set (GT Test) and the ResQ-UAV corrupted test set (Corrupted Test) constructed in this study. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e details the performance comparison of each model across the two domains; results indicate that all algorithms exhibited varying degrees of accuracy degradation within the corrupted environment.\u003c/p\u003e \u003cp\u003eTo further investigate the intrinsic mechanisms underlying this cross-domain performance disparity, we employed the RTMDet model as a case study to conduct a visual analysis of its detailed training process and statistical characteristics under both the clean source domain (GT) and the corrupted target domain (Corrupted), as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Specifically, sub-figure (a) depicts the trajectory of training accuracy improvement as iteration epochs increase. It lucidly reveals that when training on corrupted data, the ascent rate of key metrics such as \u003cspan\u003e$\u003c/span\u003emAP\u003cspan\u003e$\u003c/span\u003e and \u003cspan\u003e$\u003c/span\u003eAP_{50}\u003cspan\u003e$\u003c/span\u003e lags behind, and the ultimate peak is significantly lower than the clean baseline. This is further corroborated by the loss convergence curves in sub-figure (c), which demonstrate that the model convergence process within the corrupted environment is accompanied by higher loss values and greater oscillation amplitude.\u003c/p\u003e \u003cp\u003eConcurrently, sub-figures (b) and (d) utilize violin plots to quantify the probability distribution characteristics of performance metrics and loss values. The results indicate that the \u003cspan\u003e$\u003c/span\u003emAP\u003cspan\u003e$\u003c/span\u003e distribution for corrupted data is flatter with higher dispersion, while the loss distribution exhibits a pronounced long-tail effect. This disparity in distributional characteristics implies that the substantial volume of \"hard samples\" generated by bitstream corruption significantly exacerbates the optimization difficulty and uncertainty of the model within the feature space, thereby substantiating the fundamental cause of the quantitative performance degradation observed in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e further visualizes the differences in training dynamics of the classic two-stage detector, Faster R-CNN, between the clean source domain (GT) and the corrupted target domain (Corrupted). Observing the accuracy evolution trajectory in sub-figure (a), it is evident that although the model exhibits high stability throughout the training cycle (characterized by relatively smooth curves), the mAP metrics on the corrupted data (blue-toned curves) are consistently suppressed below the clean baseline (red-toned curves), forming a significant and persistent \"performance gap.\" This phenomenon is corroborated by the loss convergence curves in sub-figure (c); while the model is capable of converging within the corrupted environment, its loss baseline remains persistently higher than that of the clean environment, indicating that video artifacts continuously interfere with the feature matching process of the Region Proposal Network (RPN). Furthermore, the violin plots in sub-figure (b) show that the mAP distribution of Faster R-CNN is extremely compact, reflecting the low-variance characteristic of its training process; however, the center of mass of the corrupted group's distribution is shifted downward overall. Meanwhile, sub-figure (d) reveals that the loss distribution of the corrupted data possesses a longer upper tail, suggesting the presence of a substantial number of hard samples within corrupted video frames that induce high regression errors, thereby limiting the model's ultimate performance ceiling in extreme environments.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e focuses on the training characteristics of the multi-stage detection architecture, Cascade R-CNN, across different data domains. By analyzing the accuracy evolution trajectory in sub-figure (a), it is evident that although the Cascade architecture is designed to progressively enhance detection quality at high IoU thresholds via a cascade regression mechanism, the upward trend of its mAP metrics on corrupted data (blue-toned curves) significantly lags behind the clean baseline (red-toned curves). Furthermore, the performance gap between the two domains does not notably narrow as training epochs increase.\u003c/p\u003e \u003cp\u003eThis phenomenon is further corroborated by the loss convergence curves in sub-figure (c), where loss values in the corrupted environment consistently remain at a higher level. This indicates that the non-linear noise introduced by video transmission severely hinders the refined feature alignment performed by the multi-stage detection heads. Additionally, the statistical distribution plots in sub-figures (b) and (d) demonstrate that, compared to the highly concentrated performance distribution observed with clean data, the corrupted data causes an overall downward shift and increased dispersion in the mAP distribution. Simultaneously, the long-tail distribution characteristics of the loss values suggest that, even with the adoption of a cascade optimization strategy, the model continues to face significant uncertainty and optimization difficulties when confronting severe blocking artifacts and blurring artifacts.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e finally illustrates the distinct training behavior of DINO, a Transformer-based end-to-end detection paradigm, when confronted with video corruption. Analysis of the accuracy evolution trajectory in sub-figure (a) indicates that, despite DINO possessing robust global context awareness capability via its self-attention mechanism, the ascent rate and final convergence value of its mAP metrics on corrupted data (blue-toned curves) remain significantly lower than the clean baseline (red-toned curves).\u003c/p\u003e \u003cp\u003eThis characteristic is manifested uniquely in the loss convergence curves in sub-figure (c). Unlike CNN architectures, DINO's loss descent curves are extremely smooth, suggesting a relatively stable optimization process; however, a constant \"parallel gap\" persists between the clean and corrupted curves. This implies that bitstream noise induces a systematic bias in the Bipartite Matching process.\u003c/p\u003e \u003cp\u003eFurthermore, comparing the loss distributions in sub-figure (d) reveals that the loss values for clean data exhibit an extremely sharp and concentrated \"needle-like\" distribution, reflecting the Transformer's exceptionally high fitting certainty regarding clear samples. However, within the corrupted environment, the base of this distribution significantly broadens and exhibits a trend toward dispersion. This indicates that while the model is capable of processing the majority of corrupted samples, video artifacts severely impair the feature semantics of certain targets. Consequently, the model fails to effectively focus when processing these long-tail \"hard samples,\" thereby limiting any breakthrough in overall robustness.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBuilding upon the previously identified optimization difficulties induced by corrupted data, Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e and Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e further quantify the severe consequences of this training bottleneck at the inference level. The overall performance evaluation indicates that when test data is subjected to H.264 bitstream corruption, all State-of-the-Art (SOTA) models suffer from catastrophic accuracy degradation.\u003c/p\u003e \u003cp\u003eAs presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the Relative Performance Drop (RPD) predominantly falls within the high range of 29% to 36%. Specifically, Faster R-CNN exhibits an RPD as high as 36.40%, while even the relatively robust RTMDet incurs a performance loss of 29.71%. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e(a) and (b) intuitively visualize this phenomenon via histograms, where a significant disparity exists between the blue bars (representing the clean source domain) and the orange bars (representing the corrupted target domain). This confirms that models trained exclusively on clean data are incapable of effectively resisting feature space distortions inherent in video transmission, demonstrating extremely weak generalization capabilities in corrupted environments.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDetection Performance and Robustness Evaluation of Algorithms on the ResQ-UAV Benchmark\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAlgorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eGT Test\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eCorrupted Test\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c10\" namest=\"c8\"\u003e \u003cp\u003eRPD\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAP50\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAPs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAP50\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eAPs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eAP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eAP50\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eAPs\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRTMDet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.525\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.796\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.406\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.369\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.622\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.273\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e29.71%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e21.86%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e32.76%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFaster R-CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.511\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.747\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.417\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.325\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.521\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.249\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e36.40%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e30.25%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e40.29%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCascade R-CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.539\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.799\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.439\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.362\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.586\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.272\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e32.84%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e26.66%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e38.04%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDINO\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.612\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.852\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.526\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.402\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.638\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.328\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e34.31%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e25.12%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e37.64%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eA more critical finding emerges from an in-depth analysis of scale-specific metrics. By comparing the APs data in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, it is evident that the Relative Performance Drop (RPD) values for small object detection across all models (32.76%\u0026ndash;40.29%) are significantly higher than their respective declines in overall AP. This trend is visually corroborated in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e(c), which illustrates that the \"performance gap\" for small objects is the widest among all algorithms.\u003c/p\u003e \u003cp\u003eThe physical mechanism underlying this phenomenon lies in the fact that the VisDrone dataset contains a vast number of tiny targets captured from long distances, which occupy an extremely low pixel proportion. Conversely, macroblock loss and blocking artifacts induced by H.264 bitstream errors typically possess fixed spatial dimensions. When the scale of these artifacts is comparable to or even larger than that of the targets, the corrupted regions directly occlude or completely destroy the edge and texture information of small objects, causing the detector to entirely lose its feature response capability for such targets.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIn conclusion, the experimental results demonstrate that deploying models trained via traditional clean-data paradigms directly into real-world UAV links subject to communication interference poses significant risks. Existing datasets and training strategies fail to encompass the complex signal distortions induced by bitstream corruption. This limitation not only validates the universality of the cross-domain performance degradation observed in this study but also underscores the urgency and engineering value of constructing the ResQ-UAV dataset and conducting targeted robustness research.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eEffectiveness Analysis of the ResQ-UAV Dataset\u003c/h2\u003e \u003cp\u003eTo validate the effectiveness of the ResQ-UAV dataset in enhancing model robustness against interference, we substituted the original training set with the ResQ-UAV corrupted training set. The models were then retrained under identical hyperparameter settings and subsequently evaluated on the corrupted validation set. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents a detailed direct comparison of model performance pre- and post-retraining.\u003c/p\u003e \u003cp\u003eExperimental results indicate that by incorporating the ResQ-UAV dataset for fine-tuning, all evaluated models achieved significant performance recovery. Specifically, employing the Performance Gain (GAP) as a quantitative metric, the results demonstrate that GAP values for all models are positive. Notably, the AP50 for RTMDet and Faster R-CNN improved by 4.30% and 5.90%, respectively. This trend is visually corroborated in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e, where the green bars representing retraining performance surpass the corrupted baseline (yellow bars) across the vast majority of metrics.\u003c/p\u003e \u003cp\u003eIt is worth noting that the ResQ-UAV dataset not only restored overall detection accuracy but, more critically, significantly improved perception capabilities for Small Objects. For instance, Faster R-CNN achieved a gain of 1.80% in the APs metric, indicating that the dataset effectively reconstructed fine-grained features that were previously obscured by compression noise.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation of Performance Recovery and Gain (GAP) Following Retraining with the ResQ-UAV Dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"10\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c10\" colnum=\"10\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\" morerows=\"1\" rowspan=\"2\"\u003e \u003cp\u003eAlgorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c4\" namest=\"c2\"\u003e \u003cp\u003eCorrupted Test\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c7\" namest=\"c5\"\u003e \u003cp\u003eRetrain Test\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"3\" nameend=\"c10\" namest=\"c8\"\u003e \u003cp\u003eGAP\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAP50\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAPs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAP50\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eAPs\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eAP\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eAP50\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c10\"\u003e \u003cp\u003eAPs\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRTMDet\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.369\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.622\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.273\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.392\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.665\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.282\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e+\u0026thinsp;2.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e+\u0026thinsp;4.30%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e+\u0026thinsp;0.90%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFaster R-CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.325\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.521\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.249\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.336\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.580\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.267\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e+\u0026thinsp;1.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e+\u0026thinsp;5.90%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e+\u0026thinsp;1.80%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCascade R-CNN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.362\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.586\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.272\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.380\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.616\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.286\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e+\u0026thinsp;1.80%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e+\u0026thinsp;3.00%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e+\u0026thinsp;1.40%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDINO\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.402\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.638\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.328\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.413\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.673\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.314\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e+\u0026thinsp;1.10%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e+\u0026thinsp;3.50%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c10\"\u003e \u003cp\u003e-1.40%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo further validate the quality of the ResQ-UAV dataset and its stability during the optimization process, we analyzed the convergence behavior of the models during retraining. Experimental observations revealed that when training on corrupted data, both the Classification Loss and Regression Loss exhibited smooth and rapid descent trends, with a complete absence of gradient explosion or oscillation throughout the entire process. This provides compelling evidence that while the generated ResQ-UAV dataset simulates authentic bitstream corruption, it preserves adequate semantic structural information to support gradient descent and feature learning within the models. With the increase in training epochs, the validation mAP demonstrated a steady ascent and ultimately converged near the model's theoretical upper bound on the clean dataset. This further corroborates that employing ResQ-UAV for targeted domain adaptation training constitutes an effective approach to resolving detection failures induced by UAV video transmission corruption.\u003c/p\u003e \u003c/div\u003e "},{"header":"Usage Notes","content":"\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003cp\u003e \u003cb\u003eData Accessibility and Standardization.\u003c/b\u003e The ResQ-UAV dataset is now fully accessible to the academic community via a public repository, intended to accelerate research and development in robust UAV vision systems. To maximize data usability and interoperability, we have implemented strict standardization regarding the directory structure and annotation formats. The dataset adopts a hierarchical storage architecture wherein high-quality Source Images and their corresponding Bitstream-Corrupted Images share identical file naming conventions and sequence identifiers. This ensures precise indexing and seamless transition between data from different domains. Furthermore, all object annotations adhere to the standard MS COCO JSON format. This implies that researchers can directly integrate this dataset into mainstream deep learning frameworks, such as MMDetection and Detectron2, for experimentation without the need for cumbersome format conversions.\u003c/p\u003e \u003cp\u003e \u003cb\u003ePotential Application Scenarios and Extended Value.\u003c/b\u003e Beyond serving as a benchmark for robust object detection, ResQ-UAV, leveraging its unique \"pixel-level alignment\" characteristic, exhibits significant potential for exploration across broader computer vision domains:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eVideo Restoration and Enhancement\u003c/b\u003e: Given that the dataset provides pixel-level aligned \"clean-corrupted\" image pairs, it naturally constitutes an ideal training set for supervised learning tasks. Researchers can utilize pristine frames as pixel-level Ground Truth to train deep neural networks for tasks such as video De-blocking, Artifact Removal, and Super-Resolution Reconstruction, specifically targeting the complex non-linear distortions introduced by H.264 codec errors.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eRobust Multi-Object Tracking\u003c/b\u003e: Since the dataset fully preserves the temporal continuity of the original video sequences, it naturally extends the boundaries of research on temporal consistency. The availability of such sequential data enables the community to deeply analyze how Packet Loss and mosaic artifacts disrupt the continuity of motion trajectories. This opens new avenues for developing \"resilient tracking algorithms\"\u0026mdash;algorithms capable of maintaining target identity (ID) stability even when visual features are intermittently lost due to transmission failures.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eSmall Object Detection in Noisy Environments\u003c/b\u003e: Considering the prevalence of tiny objects within VisDrone scenarios, ResQ-UAV constitutes a highly challenging testbed. It not only challenges the model's feature extraction capabilities regarding small objects but also tests the limits of attention mechanisms in distinguishing foreground targets from background noise under extremely low Signal-to-Noise Ratio (SNR) conditions. This fosters the further evolution of Feature Pyramid Network (FPN) designs and context modeling techniques.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e "},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo ensure research reproducibility and foster collaborative innovation in robust UAV vision systems, the ResQ-UAV dataset and its accompanying algorithmic benchmark code have been made fully open-source to the academic community, accessible at: https://github.com/dj-dengjian/ResQ-UAV.git. This repository hosts rigorously standardized data organization scripts and MS COCO-formatted annotation files, ensuring precise alignment in terms of filenames and sequence identifiers between pristine source images and bitstream-corrupted images. By providing these core resources, we aim not only to support benchmark evaluations for robust object detection tasks but also to establish an ideal experimental platform for video de-blocking, artifact removal, super-resolution reconstruction, and resilient multi-object tracking algorithms designed to handle transmission faults. Consequently, this initiative facilitates the exploration of boundaries in computer vision technology regarding complex noisy environments and small object perception.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by the National Social Science Fund of China (No. 22\u0026amp;ZD169) and the Key project of Civil Aviation Joint Fund of National Natural Science Foundation of China (No. U2133207).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthor Jian Deng designed the experiments, generated the ResQ-UAV dataset, performed the benchmark evaluations, and wrote the original draft. Author Zeyu Liu assisted in data annotation and the implementation of the detection algorithms. Author Zihan Yu validated the experimental results. Author Honghai Zhang supervised the project, provided guidance and funding support, and reviewed and edited the paper. All authors reviewed the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ecompeting interests\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eMester, B., Frieler, K., Korup, O., Desai, B. \u0026amp; Schewe, J. Socioeconomic Predictors of Vulnerability to Flood-Induced Displacement. \u003cem\u003eNat. Commun.\u003c/em\u003e 16, (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePesonen, J. et al. Boreal Forest Fire: Uav-Collected Wildfire Detection and Smoke Segmentation Dataset. \u003cem\u003eSci. Data\u003c/em\u003e. 12, 1419 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeng, J., Zhang, H., Zhang, Y., Hua, M. \u0026amp; Sun, Y. A Method for Uav Path Planning Based On G-Maponet Reinforcement Learning. \u003cem\u003eDrones.\u003c/em\u003e 9, 871 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDeng, J., Zhang, H., Zhang, Y. \u0026amp; Sun, Y. Research On Trajectory Planning for a Limited Number of Logistics Drones (\u0026le;\u0026thinsp;3) Based On Double-Layer Fusion Gwop. \u003cem\u003eDrones.\u003c/em\u003e 9, 671 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eK. R., A. et al. Manipal-Uav Person Detection Dataset: A Step Towards Benchmarking Dataset and Algorithms for Small Object Detection. \u003cem\u003eIsprs-J. Photogramm. Remote Sens.\u003c/em\u003e 195, 77\u0026ndash;89 (2023).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSim\u0026otilde;es, D. P., Oliveira, H. C. D. \u0026amp; Pereira, D. R. Unicamp-Uav: An Open Dataset for Human Detection in Uav Imagery. \u003cem\u003eIsprs-J. Photogramm. Remote Sens.\u003c/em\u003e 231, 119\u0026ndash;136 (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobicquet, A., Sadeghian, A., Alahi, A. \u0026amp; Savarese, S. Learning Social Etiquette: Human Trajectory Understanding in Crowded Scenes. \u003cem\u003e14th European Conference on Computer Vision (ECCV\u003c/em\u003e 2016). Amsterdam, The Netherlands, 2016:549\u0026ndash;565.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eA. V., S., Sankaran, P. \u0026amp; C. V., R. Fine-Tuned Deep Models for Niche Datasets \u0026mdash; People Detection in Uav Building Images to Aid Rescue Operations. \u003cem\u003eInt. J. Appl. Earth Obs. Geoinf.\u003c/em\u003e 145, 104985 (2025).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWiegand, T., Sullivan, G. J., Bjontegaard, G. \u0026amp; Luthra, A. Overview of the H.264/Avc Video Coding Standard. \u003cem\u003eIeee Trans. Circuits Syst. Video Technol.\u003c/em\u003e 13, 560\u0026ndash;576 (2003).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRen, S., He, K., Girshick, R. \u0026amp; Sun, J. Faster R-Cnn: Towards Real-Time Object Detection with Region Proposal Networks. In: Cortes C, Lawrence ND, Lee DD, Sugiyama M, Garnett R, eds. \u003cem\u003eADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 28 (NIPS 2015)\u003c/em\u003e. 29th Annual Conference on Neural Information Processing Systems (NIPS), 2015.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCai, Z. W., Vasconcelos, N. \u0026amp; IEEE. Cascade R-Cnn: Delving Into High Quality Object Detection. \u003cem\u003e2018 IEEE/CVF CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)\u003c/em\u003e. 31st IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018:6154\u0026ndash;6162.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLyu, C. et al. Rtmdet: An Empirical Study of Designing Real-Time Object Detectors. \u003cem\u003eArxiv\u003c/em\u003e. abs/2212.07784, (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, H. et al. Dino: Detr with Improved Denoising Anchor Boxes for End-to-End Object Detection. \u003cem\u003eInternational Conference on Learning Representations (ICLR)\u003c/em\u003e, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu, X., Li, W., Hong, D., Tao, R. \u0026amp; Du, Q. Deep Learning for Unmanned Aerial Vehicle-Based Object Detection and Tracking: A Survey. \u003cem\u003eIeee Geosci. Remote Sens. Mag.\u003c/em\u003e 10, 91\u0026ndash;124 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu, P. et al. Detection and Tracking Meet Drones Challenge. \u003cem\u003eIeee Trans. Pattern Anal. Mach. Intell.\u003c/em\u003e 44, 7380\u0026ndash;7399 (2022).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGui-Song Xia, X. B. J. D. \u0026amp; Jiebo Luo, M. D. M. P. Dota: A Large-Scale Dataset for Object Detection in Aerial Images. 2018 \u003cem\u003eIeee/Cvf Conference On Computer Vision and Pattern Recognition\u003c/em\u003e. (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu, X., Gong, Y., Jiang, N., Ye, Q. \u0026amp; Han, Z. Scale Match for Tiny Person Detection. 2020 \u003cem\u003eIeee Winter Conference On Applications of Computer Vision (Wacv)\u003c/em\u003e. 1246\u0026ndash;1254 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuszak, A. \u0026amp; Imre, S. Analysing Gop Structure and Packet Loss Effects On Error Propagation in Mpeg-4 Video Streams. \u003cem\u003e4th International Symposium on Communications, Control and Signal Processing\u003c/em\u003e: IEEE, 2010:1\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, T. et al. Bitstream-Corrupted Video Recovery: A Novel Benchmark Dataset and Method. \u003cem\u003eAdvances in Neural Information Processing Systems (NeurIPS)\u003c/em\u003e, 2023.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGhasempour, M. \u0026amp; Ghanbari, M. A Low Complexity System for Multiple Data Embedding Into H.264 Coded Video Bit-Stream. \u003cem\u003eIeee Trans. Circuits Syst. Video Technol.\u003c/em\u003e 30, 4009\u0026ndash;4019 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ., L. et al. A Comprehensive Benchmark for Single Image Compression Artifact Reduction. \u003cem\u003eIeee Trans. Image Process.\u003c/em\u003e 29, 7845\u0026ndash;7860 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWiegand, T., Schwarz, H., Joch, A., Kossentini, F. \u0026amp; Sullivan, G. J. Rate-Constrained Coder Control and Comparison of Video Coding Standards. \u003cem\u003eIeee Trans. Circuits Syst. Video Technol.\u003c/em\u003e 13, 688\u0026ndash;703 (2003).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStockhammer, T., Hannuksela, M. M. \u0026amp; Wiegand, T. H.264/Avc in Wireless Environments. \u003cem\u003eIeee Trans. Circuits Syst. Video Technol.\u003c/em\u003e 13, 657\u0026ndash;673 (2003).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhawaja, W., Guvenc, I., Matolak, D. W., Fiebig, U. \u0026amp; Schneckenburger, N. A Survey of Air-to-Ground Propagation Channel Modeling for Unmanned Aerial Vehicles. \u003cem\u003eIeee Commun. Surv. Tutor.\u003c/em\u003e 21, 2361\u0026ndash;2391 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKurutepe, E., Civanlar, M. R. \u0026amp; Tekalp, A. M. Client-Driven Selective Streaming of Multiview Video for Interactive 3Dtv. \u003cem\u003eIeee Trans. Circuits Syst. Video Technol.\u003c/em\u003e 17, 1558\u0026ndash;1565 (2007).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeshadrinathan, K. \u0026amp; Bovik, A. C. Motion Tuned Spatio-Temporal Quality Assessment of Natural Videos. \u003cem\u003eIeee Trans. Image Process.\u003c/em\u003e 19, 335\u0026ndash;350 (2010).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang, S. et al. Widerperson: A Diverse Dataset for Dense Pedestrian Detection in the Wild. \u003cem\u003eIeee Trans. Multimedia\u003c/em\u003e. 22, 380\u0026ndash;393 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSakaridis, C., Dai, D. \u0026amp; Van Gool, L. Semantic Foggy Scene Understanding with Synthetic Data. \u003cem\u003eInt. J. Comput. Vis.\u003c/em\u003e 126, 973\u0026ndash;992 (2018).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTeney, D., Liu, L., van den Hengel, A. \u0026amp; IEEE. Graph-Structured Representations for Visual Question Answering. \u003cem\u003e30TH IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR 2017)\u003c/em\u003e. 30th IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2017:3233\u0026ndash;3241.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin, T. et al. Microsoft Coco: Common Objects in Context. In: Fleet D, Pajdla T, Schiele B, Tuytelaars T, eds. \u003cem\u003eCOMPUTER VISION - ECCV 2014, PT V\u003c/em\u003e. 13th European Conference on Computer Vision (ECCV), 2014:740\u0026ndash;755.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKai, C. et al. Mmdetection: Open Mmlab Detection Toolbox and Benchmark [Arxiv]. \u003cem\u003eArxiv\u003c/em\u003e. 13 (2019).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCampillos-Llanos, L. et al. Transformer-Based Relation Extraction and Concept Normalization Using an Annotated Clinical Trials Corpus. \u003cem\u003eSci. Data\u003c/em\u003e. (2026).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQian, N. On the Momentum Term in Gradient Descent Learning Algorithms. \u003cem\u003eNeural Netw.\u003c/em\u003e 12, 145\u0026ndash;151 (1999).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun, S., Cao, Z., Zhu, H. \u0026amp; Zhao, J. A Survey of Optimization Methods From a Machine Learning Perspective. \u003cem\u003eIeee T. Cybern.\u003c/em\u003e 50, 3668\u0026ndash;3681 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang, L. \u0026amp; Shami, A. On Hyperparameter Optimization of Machine Learning Algorithms: Theory and Practice. \u003cem\u003eNeurocomputing\u003c/em\u003e. 415, 295\u0026ndash;316 (2020).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan, M., Pang, R. \u0026amp; Le, Q. V. Efficientdet: Scalable and Efficient Object Detection. 2020 \u003cem\u003eIEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)\u003c/em\u003e, 2020.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu, J. et al. Benchmarking Object Detection Robustness Against Real-World Corruptions. \u003cem\u003eInt. J. Comput. Vis.\u003c/em\u003e 132, 4398\u0026ndash;4416 (2024).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8741863/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8741863/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThis paper presents and open-sources ResQ-UAV, a dataset designed for person detection in corrupted videos from the perspective of Unmanned Aerial Vehicles (UAVs). Generated using H.264 bitstream corruption techniques, the dataset aims to simulate realistic non-linear signal impairments encountered in actual image transmission links. By mimicking authentic communication link failures, the images exhibit highly complex degradation characteristics, posing significant challenges for detection algorithms operating under communication-constrained conditions. ResQ-UAV comprises 16,414 frames derived from 32 video sequences, featuring a total of 282,095 meticulously annotated bounding boxes for persons. It encompasses diverse complex scenarios, including urban arterials and suburban areas, across various lighting conditions (day and night). Serving as a dedicated benchmark for evaluating video transmission robustness, this dataset significantly enriches the data resources available for UAV vision in complex transmission environments. Benchmarking results based on various state-of-the-art object detection algorithms demonstrate that ResQ-UAV establishes a rigorous performance baseline for multi-object detection tasks within corrupted video environments. The dataset is poised to provide a reliable data foundation and verification platform for the research and application of critical technologies in disaster rescue operations. Data and code are available at: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/dj-dengjian/ResQ-UAV.git\u003c/span\u003e\u003cspan address=\"https://github.com/dj-dengjian/ResQ-UAV.git\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/p\u003e","manuscriptTitle":"ResQ-UAV: A Novel Dataset Supporting Robust Recognition for Future Drone Rescue Missions","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-23 05:35:02","doi":"10.21203/rs.3.rs-8741863/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"47702003-82de-4d4b-9810-ae5af1c11abf","owner":[],"postedDate":"February 23rd, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-04-23T15:55:23+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-23 05:35:02","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8741863","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8741863","identity":"rs-8741863","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}