RPV11K: A Benchmark for Joint Product-Vacancy Detection in Retail Scenarios

RPV11K: A Benchmark for Joint Product-Vacancy Detection in Retail Scenarios | 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 RPV11K: A Benchmark for Joint Product-Vacancy Detection in Retail Scenarios Bidong Chen, Lingui Li, Yuanda Lin, Xu yang, Sio Kei Im, RuiPedro Paiva, and 1 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6428418/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 In intelligent retail, accurate detection of densely arranged products and shelf vacancies in unstructured environments remains a critical challenge. This paper introduces RPV11K, a large-scale benchmark dataset (11,743 im-ages, 1.87M annotations) designed for joint product and vacancy detection in real-world retail scenarios. We develop a novel edge-deployable framework, DetectRPV, integrating lightweight YOLO variants and systematic data augmentation. Experimental results on the Jetson edge platform show that YOLOv1-medium achieves the best performance with mAP50 of 80.6% and mAP of 55.1%, while maintaining an inference speed of 52ms. RPV11K, the first dataset to explicitly model both product and vacancy categories under dense occlusion and diverse lighting, provides a rigorous evaluation benchmark for automated shelf management systems. Our work bridges the gap between academic research and industrial deployment by establishing a benchmark for real-time shelf monitor-ing with metrics relevant to automated retail operations. The code and data will be made available on https://github.com/Cy1oong/RPV11K . Physical sciences/Engineering/Electrical and electronic engineering Physical sciences/Mathematics and computing Physical sciences/Mathematics and computing/Applied mathematics Physical sciences/Mathematics and computing/Computational science Physical sciences/Mathematics and computing/Computer science Physical sciences/Mathematics and computing/Information technology Physical sciences/Mathematics and computing/Pure mathematics Physical sciences/Mathematics and computing/Scientific data Physical sciences/Mathematics and computing/Software Physical sciences/Mathematics and computing/Statistics Intelligent retail vacancy detection product detection edge deployment shelf monitoring Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1 Introduction In the era of digital transformation, intelligent retail is revolutionizing traditional business models through the integration of Artificial Intelligence (AI), robotics, and Internet of Things (IoT) technologies [1, 2]. This technological convergence is fundamentally reshaping shopping experiences and streamlining operational processes [3, 4]. Among various retail operations, shelf management stands as a critical component that directly impacts business success through its operational efficiency and accuracy. Conventional shelf management methods depend predominantly on manual inspection, wherein personnel are required to physically assess aisles to ascertain product availability, location, and inventory levels [5, 6]. This labor-intensive approach is costly and subject to human error, resulting in challenges such as vacancy, misplaced products, and inefficient inventory management. The repercussions of merchandise shortages on customer satisfaction and brand reputation are substantial when consumers encounter such circumstances [7]. The advent of unmanned retail establishments represents a paradigm shift in the domain of retail digitalization. These intelligent spaces employ sophisticated technology, including artificial intelligence (AI) powered computer vision systems, robotics, and the internet of things (IoT), to automate various operational aspects [4, 8]. Specifically, computer vision-based monitoring systems play a pivotal role in shelf management, vacancy detection, and automated replenishment. These systems possess the capability to initiate robotic replenishment upon detecting shelf vacancies, thereby establishing a completely automated workflow from the detection of vacancies to their resolution [8]. This automation substantially improves both operational efficiency and the overall shopping experience in retail environments. However, the development of effective product and vacancy detection systems for real-world retail environments poses significant technical challenges. The unstructured nature of supermarket shelves, marked by dense product arrangements, diverse categories, and variable lighting conditions, gives rise to complex scenarios for detection algorithms. The dense arrangement of products gives rise to occlusion issues, while the diversity of products in terms of shape, size, and packaging increases the complexity of detection. Furthermore, varying illumination conditions, ranging from intense spotlights to dim aisles, have been shown to significantly affect the quality of visual information captured by monitoring systems. To address these challenges and advance the research into intelligent retail detection systems, our contributions are as follows: RPV11K Dataset: We introduce a shelf product and vacancy detection dataset containing 11,743 images and more than 1.87 million annotations, covering different retail scenarios, laying a solid foundation for future research. Extensive Benchmarking: We conduct thorough evaluations using a series of lightweight object detection models at different scales, establishing baseline performance metrics for edge computing deployment. The benchmark provides valuable insights into model performance under real-world retail constraints. Edge Computing Analysis: We explore model deployment on the Jetson-nano platform, providing preliminary performance analysis and practical observations for real-time retail detection systems. These findings may help bridge academic research and industrial applications. The remainder of this paper is organized as follows: Section 2 reviews related work in object detection and retail applications. Section 3 details the proposed methodology. Section 4 presents experimental results and analysis. Finally, Section 5 concludes the paper with future directions. 2 Related work 2.1 Intelligent retail Intelligent retail represents a significant advancement in the retail industry through the integration of advanced technologies to enhance customer experience and operational efficiency. Modern retail automation largely relies on artificial intelligence, IoT, and robotics technologies. Wang et al. [9] introduced an IoT-based inventory management system utilizing micro-sensors embedded in store shelves for product movement detection and inventory monitoring. While innovative, this approach faces practical limitations regarding deployment costs and maintenance requirements. In unmanned retail research, Zhang et al. [10] explored computer vision-based product recognition and pricing systems. Their approach employs deep learning algorithms for feature extraction and matching from product packaging. However, the system’s performance tends to decrease under challenging conditions, such as similar product packaging, partial occlusions, and varying lighting conditions. While considerable research has addressed shopping process optimization, studies on automated shelf management, particularly vacancy detection, remain limited. Although Chen [11] and Šikić F [12] recognized the importance of shelf management, they have yet to propose effective solutions for product vacancy detection. This suggests potential opportunities for developing robust shelf management solutions. 2.2 Existing dataset The availability of data plays a foundational role in the advancement of intelligent retail technology, particularly through the utilization of open-source datasets. Early attempts to identify retail products had obvious flaws, particularly in terms of scale and scope. The LabelMe Food dataset [13], a pioneering achievement in image annotation research, contains a limited number of food product images with crowdsourced annotations, thereby restricting its applicability in retail scenarios. A similar limitation is exhibited by the Food101 dataset [14], which, despite its substantial size of 101,000 images across 101 categories, focuses exclusively on food items in controlled environments, thereby failing to capture the complexity of real retail settings. As the field develops, datasets of more complex scenarios gradually become available. The MS COCO-Stuff dataset [15], though not retail-specific, contributed valuable insights through its diverse object categories in various contexts. The Open Images Dataset [16] further expanded this scope, incorporating retail products among its numerous categories. Recent advancements have led to a shift towards real-world retail environments. The RP2K dataset [17] serves as a prime example. Although it is mainly centered around ecommerce applications, it offers valuable perspectives on the visual features of products in actual shopping scenarios. More recent contributions include the AiProducts-Challenge dataset [18,19], containing approximately 3 million images across 50,000 SKU-level categories, marking a significant advancement in scale. However, it presents challenges such as noisy data, unbalanced distribution, and annotation limitations. The SKU-110k dataset [20] specifically addresses dense product detection, while the RPC dataset [21] contributes 83,739 standardized images across 200 SKU products, captured through systematic photography methods. Our research differs fundamentally by focusing on shelf vacancy detection, introducing unique challenges beyond traditional product recognition. The RPV11K dataset specifically addresses complex retail scenarios including product occlusion and dense placement, while uniquely incorporating both product and vacant area detection capabilities. This dual-category approach presents enhanced technical challenges compared to conventional single-category vacancy detection methods. To better demonstrate the uniqueness of our proposed RPV11K dataset, in addition to the datasets described above, we also make a comprehensive comparison with a variety of existing open-source datasets, as shown in Table 1. The comparison covers several key aspects, including dataset scale, annotation density, and detection capabilities. Our dataset shows significant advantages in terms of the number of images, bounding box density, and the integration of the most prominent product and vacancy detection features. Table 1 . Key attributes of relevant benchmarks Name Time Images quantity Bbox quantity Box/image Category High Densely Vacancy Retail PASCAL VOC[39] 2012 123287 896782 7.27 80 × × × Grocery Shelves[4] 2015 354 13184 37.24 10 × × ✓ COCO2017[40] 2017 123287 52090 2.42 20 × × × Dota-v2.0[41] 2021 11268 1793658 159.18 18 ✓ × × SKU-110K[20] 2019 2021 11762 1730996 147.17 1 ✓ × ✓ Locount[42] 50394 809659 16.07 140 × × ✓ PRV12K(ours) 2025 11743 1874077 159.59 2 ✓ ✓ ✓ Bbox quantity is short for bounding box, Box/image indicates the average number of bounding boxes per image; ✓ indicates that it has this attribute; × indicates that there is no such attribute. The original SKU-110K has a portion of the image that is damaged and of poor quality. 2.3 Retail Object Detection Deep learning-based methods have demonstrated significant advancements in product object detection. The YOLO series has established a prominent position in industrial applications, balancing real-time performance with detection accuracy. The initial YOLOv1, introduced by Redmon J et al. [22], innovatively approached object detection as a regression problem, enabling direct object localization and classification from global image features, thus substantially improving detection speed. Subsequent YOLO iterations brought notable improvements. YOLOv5 [23] incorporated the CSPNet (Cross Stage Partial Network) structure to enhance feature propagation, achieving higher accuracy while maintaining detection efficiency. YOLOv6 [24] streamlined the network architecture through RepVGG structure, converting multi-branch features to single-branch structures during inference, thereby optimizing computational efficiency. YOLOv8 [34] further advanced performance through improved loss functions and data augmentation strategies, including enhanced CIoU loss for better multi-scale object detection. Despite these advances, significant challenges persist in complex retail environments. In supermarket shelf scenarios characterized by densely arranged products and varying lighting conditions, these algorithms face performance limitations. Zhang et al. [25] identified that YOLO-based models often struggle with detection accuracy when encountering densely packed or partially occluded products. 2.4 Vacancy Position Detection Vacancy position detection represents a critical aspect of shelf management, yet research in this domain remains relatively nascent. Current approaches can be categorized into three main directions: image analysis, object detection algorithms, and multimodal data fusion. Image analysis approaches, exemplified by Chen et al. [26], employ background subtraction techniques to construct shelf background models and identify empty regions through comparative analysis. While innovative, these methods demonstrate significant sensitivity to lighting variations and often produce false detections in complex retail environments. Object detection-based approaches offer an alternative methodology. Arora D et al. [27] implemented Faster R-CNN to detect shelf products and identify vacancies through layout comparison. However, real-world challenges including product occlusion, packaging similarities, and varying placement angles significantly impact detection reliability. While deep learning algorithms such as YOLO series, RT-DETR [28], SSD [29], and RetinaNet [30] have shown promising results in product void detection, they face persistent challenges in intelligent retail applications. These challenges primarily stem from the unstructured nature and environmental complexity of retail shelves, manifesting in: Limited detection accuracy in dense product arrangements; Insufficient real-time performance in complex scenarios; Reduced reliability under varying lighting conditions; Difficulty in handling partial occlusions. The limitations of existing approaches underscore the necessity for more effective shelf vacancy detection methods. To address the challenges of complex retail environments and overcome the constraints of current solutions, this paper proposes an enhanced detection framework that integrates optimized object detection architecture with retail-specific features. 3 Method 3.1 RPV11K dataset The SKU-110K dataset [20] consists of 11,762 high-density retail scene images containing approximately 1.73 million annotated bounding boxes. The dataset is partitioned into 8,233 training images, 588 validation images, and 2,941 test images, all captured under standardized conditions with consistent device specifications and resolution. To enhance the dataset’s robustness and practical applicability, we implemented comprehensive data augmentation through random sampling techniques to address the limited lighting variations in the original dataset. The augmentation process was supplemented by incorporating additional images from complex retail scenarios, effectively expanding the environmental diversity of the training data. Furthermore, we refined the dataset composition to better align with current retail markets. This refinement involved removing product categories uncommon in chinese retail environments while substantially supplementing the dataset with prevalent shelf products. These modifications were specifically designed to enhance the dataset’s relevance to practical application scenarios in contemporary retail settings. In Fig.1, the RPV11K dataset systematically demonstrates the multi-dimensional characteristics of retail shelf environments through diverse perspective captures. Specifically, the dataset incorporates two critical viewing angles: the merchandise wall perspective, which provides a comprehensive macroscopic visualization of shelf layout configurations, and the front view perspective provides detailed visualization of product arrangement patterns on the shelf facade. This view captures a diverse array of products with distinctive chromatic characteristics. The side-angle perspective delineates the lateral structural composition of shelves and the hierarchical organization of product placement. The dual-perspective imaging upward and downward views enables comprehensive shelf analysis, capturing both detailed upper tier products and overall distribution patterns. Additionally, the dataset incorporates random placement scenarios, authentically simulating the stochastic product arrangements that occur during routine retail operations such as inventory replenishment and promotional activities. This feature enhances both data authenticity and complexity. Regarding product density characteristics, the dataset shelves are populated with varied consumer goods categories including beverages, food items, and daily necessities. The high-density product arrangement accurately represents the characteristic spatial constraints of contemporary retail environments. In the context of vacancy detection, these diverse shelf scenarios and product placement variations provide comprehensive training materials. The multi-perspective shelf imagery and random placement situations effectively simulate real-world vacancy occurrences, facilitating the development and validation of detection algorithms. This robust data foundation ensures reliable technical support for practical applications in unmanned retail environments. As illustrated in Fig.2 (a), the quantitative relationship between product detection boxes and vacancy detection boxes exhibits a ratio of 12:1, indicating a predominant presence of product-related annotations relative to vacancy-related annotations within the dataset. Fig.2 (b) delineates the spatial distribution patterns of both product and vacancy detection boxes across individual images. Statistical analysis reveals that while products constitute the primary detection targets, the dataset encompasses partial vacancy position detection, demonstrating its comprehensive scope in addressing both product recognition and shelf vacancy monitoring. Heat map visualizations in Fig.2 (c) and Fig.2 (d) illustrate the spatial density distribution of product instances and shelf vacancies across the dataset. The spatial density analysis reveals a centrally concentrated distribution of product instances, with density gradually decreasing towards peripheral regions. In contrast, vacancy category detection boxes display distinct spatial distribution patterns, exhibiting an overall density degradation from central to peripheral regions, while maintaining localized high-density clusters. 3.2 Evaluation Metrics We evaluate detection performance using standard metrics: Average Precision (AP) for single-class and mean Average Precision (mAP) for multi-class evaluation [20]. The metrics are defined as: Where TP , FP , FN represent true positives, false positives, and false negatives respectively, and N is the total number of classes. 4 Experiments 4.1 Experimental parameter settings We employed different hyperparameter settings based on model scales. For nano and small-scale YOLO models, we set the weight decay parameter to 1e-2, with both initial and final learning rates maintained at 1e-3 and incorporated a Dropout ratio of 0.1. For medium-scale YOLO models, we increased the weight decay to 5e-2 while decreasing both initial and final learning rates to 1e-4 to ensure stable training convergence. Training was conducted on a server equipped with Intel(R) Xeon(R) Gold 6330 CPU and four NVIDIA A5000 GPUs. Data augmentation follows YOLOv11 guidelines to maintain scale invariance. The augmentation pipeline comprises Mosaic (4-image composition with scale range 0.5-1.5), random horizontal flipping (probability 0.5), and random erasure (10-30% of image area with aspect ratio 0.3-3.0). For optimization, nano-scale models (≤3M params) employ a higher weight decay (1e-2) to mitigate overfitting on small datasets, while medium models (20-60M params) adopt lower learning rates (1e-4) to stabilize gradient propagation through deeper networks. All model testing and performance evaluation were carried out exclusively on the Jetson Orin Nano Super platform (4GB RAM, 256GB storage) running Ubuntu 24.10. We adopte COCO-style metrics [31], including AP50 (average precision at IoU=50%), AP (averaged over IoU thresholds from 50% to 95% with a step of 5%), and mean Average Precision (mAP) for multi-class performance. Additionally, we utilized Precision, Recall, and F1-Score to assess detection capabilities and balance at different thresholds, while model operational efficiency was evaluated through Frames Per Second (FPS) and GFLOPs. Fig.3 outlines our DetectRPV framework. The annotation pipeline (Fig.3 a) begins with manual labeling using LabelMe [32], followed by format conversion to YOLO annotations for both product instances and shelf vacancies. The RPV11K dataset is partitioned into training, validation, and test sets following the established protocol of SKU-110K. We apply lossless compression techniques to reduce image storage requirements while preserving detection-critical features, optimizing inference speed without accuracy loss. We evaluate multiple YOLO variants on the processed dataset and deploy the optimized models on Jetson platform for real-time retail testing (Fig. 3 b). 4.2 Results Analysis Drawing upon the experimental results presented in Fig.4, we carried out a comprehensive performance assessment of various YOLO variants on the RPV11K dataset. In terms of latency-accuracy trade-off, YOLOv11 achieved the highest mAP value (approximately 55%) with a latency of 20ms, while YOLOv6 demonstrated balanced performance in terms of latency and accuracy. Regarding the comparison between model size and accuracy, YOLOv11 exhibited superior detection performance with a parameter configuration of 20.03M, validating the effectiveness of its architecture design. The training convergence curves showed that YOLOv11-m exhibited faster convergence rates and higher final accuracy compared to YOLOv9-s and YOLOv8-n. These experimental results provide important reference data for model selection in different application scenarios, and each model shows its applicability under different latency and accuracy requirements. Table 2 . Performance comparison of different-scale models on Jetson edge computing platform, ‘-’ denotes metrics unavailable due to edge device resource limitations. Models Scales Precision Recall mAP mAP 50 F1 latency FPS Params GFLOPs YOLOv5[23] nano small medium 0.781 0.710 0.514 0.778 0.744 14.1 71 2.50 7.1 0.792 0.730 0.534 0.794 0.760 24.8 40 7.82 18.7 0.794 0.743 0.546 0.804 0.768 48.5 21 22.13 52.5 YOLOv6[24] nano Small medium 0.781 0.710 0.514 0.777 0.744 13.0 77 4.23 11.8 0.789 0.732 0.531 0.793 0.759 33.55 28 15.98 42.8 0.788 0.741 0.542 0.799 0.764 - - 51.25 158.3 YOLOv8[33] nano small medium 0.787 0.714 0.520 0.782 0.749 15.6 64 3.01 8.1 0.785 0.719 0.519 0.783 0.751 28.3 35 9.82 23.3 0.791 0.744 0.546 0.803 0.767 54.8 18.25 23.22 67.4 YOLOv9[34] tiny small medium 0.785 0.712 0.520 0.782 0.747 23.3 43 1.76 6.4 0.792 0.738 0.542 0.800 0.764 35.7 28 6.19 22.1 0.789 0.737 0.536 0.796 0.762 57.0 18 20.01 76.5 YOLOv10[35] nano small medium 0.781 0.705 0.514 0.776 0.741 17.9 56 2.70 8.2 0.787 0.724 0.533 0.790 0.754 31 32 8.03 24.4 0.790 0.737 0.544 0.798 0.763 54 19 16.48 63.4 YOLOv11[36] nano small medium 0.782 0.713 0.516 0.779 0.746 17.3 58 2.58 6.3 0.793 0.730 0.534 0.795 0.760 30.0 33 9.41 21.3 0.797 0.744 0.551 0.806 0.770 52.0 19 20.03 67.7 RT-DETR[37] - 0.763 0.689 0.487 0.737 0.724 - - 63.18 103.4 Table 2 presents the comparative evaluation of YOLO variants and RT-DETR on RPV11K. YOLOv11-medium achieves superior performance across all metrics, with precision of 0.797, recall of 0.744, F1-score of 0.770, and mAP50 of 0.806. YOLOv8-medium and YOLOv9-medium show competitive results, while RT-DETR-large yields relatively lower precision (0.763). Notably, lightweight models like YOLOv9-tiny and YOLOv8-nano demonstrate strong recall performance (0.712 and 0.714 respectively). Table 3 summarizes the benchmark experiments presented in Table 2, covering the performance of models of different scales. In product detection, YOLOv9-s and YOLOv11-m demonstrate outstanding performance with remarkable precision, recall, and F1-scores, while YOLOv8-n is slightly less remarkable. Regarding product vacancy detection, the precision and recall of all models are relatively low, indicating that there is substantial room for improvement, which may be attributed to the challenging nature of vacancy features. In terms of average precision, YOLOv11-m performs best in product detection. Overall, model optimization is necessary to enhance the vacancy detection ability and contribute to the optimization of retail management. In terms of model complexity, the RT-DETR model has the largest parameter scale and the highest computational load, while the YOLOv5 nano model has the smallest parameter scale and computational load. The YOLOv11 medium model achieves a good balance between performance and complexity, showing great potential. Other models also have their own advantages in different metrics. Thus, they can be selected according to specific requirements. In the future, exploring the optimization of the model structure can be considered to improve performance while reducing complexity. Fig.5 shows the results of qualitative testing of the RPV11K dataset through the DetectRPV process. We observed a series of supermarket shelf images, which were filled with various commodities, including food, daily necessities, etc. The commodities were rich and varied, colorful, and neatly arranged. Through the DetectRPV test process, the location of the commodities in the image was clearly marked, and each commodity was circled with a different color box, which intuitively presented the model's detection effect on the commodity. From the test results, the YOLOv11 benchmark model can accurately identify commodities and detect vacant locations in complex scenes such as supermarket shelves, but there is still a lot of room for improvement. It is worthwhile for us to continue to use RPV11K as a benchmark to study models with better performance. Table 3 . Performance benchmark models of different sizes on RPV11K data Models Class Precision Recall mAP mAP 50 F1 YOLOv8-n Product 0.890 0.863 0.593 0.917 0.876 Vacancy 0.685 0.564 0.447 0.648 0.619 YOLOv9-s Product 0.897 0.886 0.613 0.932 0.891 Vacancy 0.686 0.589 0.470 0.669 0.636 YOLOv11-m Product 0.897 0.900 0.622 0.938 0.898 Vacancy 0.698 0.589 0.480 0.674 0.639 5 Conclusion Current object detectors, while successful in standard benchmarks, face significant challenges in dense retail environments. To address this gap, we introduce RPV11K - a new benchmark dataset of retail shelf images with precise annotations for both product detection and vacancy locations. Our work makes two key contributions: (1) establishing a challenging retail-focused dataset that reflects real-world complexity, and (2) conducting comprehensive evaluations of lightweight detectors suitable for edge deployment in shelf management robots. Our experiments on the Jetson-nano platform demonstrate that even state-of-the-art models like YOLOv11 achieve a limited performance, indicating substantial room for improvement in dense retail scene understanding. The RPV11K dataset, as the first large-scale benchmark focusing on retail product and shelf vacancy detection, provides a reliable evaluation framework for developing and accessing automated retail solutions, facilitating technological advancement in this domain. Declarations Acknowledgments . This work was supported by Foundation for Science and Technology (FCT) through national funds (UIDB/00326/2025, UIDP/00326/2025) and Macao Polytechnic University (Grant No. RP/FCA-04/2022). Declaration of Competing Interest. All authors declare no conflicts of interest; All authors read and approved the final manuscript; No historical conflicts of interest; No conflicts of interest in the past 3 years. Data availability. The data is constructed by ourselves, and there has been no conflict of interest in history. The experimental data in the paper will be provided upon request. If necessary, you can contact corresponding author: (Correspondence: L.L, [email protected] ; Y.W, [email protected] ;) Credit authorship contribution statement. All authors read, were informed of, and approved the final manuscript. Contributions are as follows: Conceptualization,B.C,L.L,Y.L; methodology,B.C,L.L; software,B.C,L.L validation,B.C, formal analysis,B.C,Y.W; investigation,B.C,L.L,Y.L; resources,B.C,L.L data curation,B.C,L.L; writing—original draft preparation,B.C; writing—review and editing,B.C,L.L; visualization,B.C,L.L; supervision,X.Y,S.I,R.P,Y.W; project administration,B.C,X.Y,S.I,R.P,Y.W; funding acquisition,X.Y,R.P,Y.W; All authors have read and agreed to the published version of the manuscript. Disclosure of Interests. The authors declare no conflict of interest. References Gauri D K, Jindal R P, Ratchford B, et al. Evolution of retail formats: Past, present, and future[J]. Journal of Retailing, 2021, 97(1): 42-61. Har L L, Rashid U K, Te Chuan L, et al. Revolution of retail industry: from perspective of retail 1.0 to 4.0[J]. Procedia Computer Science, 2022, 200: 1615-1625. Rane N. Integrating leading-edge artificial intelligence (AI), internet of things (IOT), and big data technologies for smart and sustainable architecture, engineering and construction (AEC) industry: Challenges and future directions[J]. Engineering and Construction (AEC) Industry: Challenges and Future Directions (September 24, 2023), 2023. Wolniak R, Stecuła K, Aydın B. Digital Transformation of Grocery In-Store Shopping-Scanners, Artificial Intelligence, Augmented Reality and Beyond: A Review[J]. Foods, 2024, 13(18): 2948. H¨ubner, A. (2011). Retail Category Management: Decision Support Systems for Assortment, Shelf Space, Inventory and Price Planning. Springer Science & Business Media. Sajadi S J, Ahmadi A. An integrated optimization model and meta-heuristics for assortment planning, shelf space allocation, and inventory management of perishable products: a real application[J]. Plos one, 2022,17(3): e0264186. Kollock P. The production of trust in online markets[J]. Advances in group processes, 1999, 16(1): 99-123. Vermesan O, Br¨oring A, Tragos E, et al. Internet of robotic things–converging sensing/actuating, hyperconnectivity, artificial intelligence and IoT platforms[M]//Cognitive hyperconnected digital transformation. River Publishers, 2022: 97-155. Wang Y C, Yang C C. 3S-cart: a lightweight, interactive sensor-based cart for smart shopping in supermarkets[J]. IEEE Sensors Journal, 2016,16(17): 6774-6781. Zhang H, Li D, Ji Y, et al. Toward new retail: A benchmark dataset for smart unmanned vending machines[J]. IEEE Transactions on Industrial informatics, 2019, 16(12): 7722-7731. Chen J, Wang S L, Lin H L. Out-of-stock detection based on deep learning[C]//Intelligent Computing Theories and Application: 15th In-ternational Conference, ICIC 2019, Nanchang, China, August 3–6, 2019, Proceedings, Part I 15. Springer International Publishing, 2019: 228-237. ˇSiki´c F, Kalafati´c Z, Subaˇsi´c M, et al. Enhanced Out-of-Stock Detection in Retail Shelf Images Based on Deep Learning[J]. Sensors, 2024, 24(2):693. Russell B C, Torralba A, Murphy K P, et al. LabelMe: a database and web-based tool for image annotation[J]. International journal of computer vision, 2008, 77: 157-173. Bossard L, Guillaumin M, Van Gool L. Food-101–mining discriminative components with random forests[C]//Computer vision–ECCV 2014:13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13. Springer International Publishing, 2014: 446-461. Caesar H, Uijlings J, Ferrari V. Coco-stuff: Thing and stuff classes in context[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 1209-1218. Krasin I, Duerig T, Alldrin N, et al. Openimages: A public dataset for large-scale multi-label and multi-class image classification[J]. Dataset available from https://github. com/openimages, 2017, 2(3): 18. Peng J, Xiao C, Li Y. RP2K: A large-scale retail product dataset for fine-grained image classification[J]. arXiv preprint arXiv:2006.12634, 2020. Liang Y, Zhang W. Winner Solution for AliProducts Challenge: Large-scale Product Recognition[J]. Cheng L, Zhou X, Zhao L, et al. Weakly supervised learning with side information for noisy labeled images[C]//Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXX 16. Springer International Publishing, 2020: 306-321. E. Goldman, R. Herzig, A. Nachman, M. Salzmann, and D. Freedman. Precise detection in densely packed scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5227-5236, 2019. Wei X S, Cui Q, Yang L, et al. RPC: A large-scale retail product checkout dataset[J]. arXiv preprint arXiv:1901.07249, 2019. Redmon J. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016. Jocher G, Chaurasia A, Stoken A, et al. ultralytics/yolov5: v7. 0-yolov5 sota realtime instance segmentation[J]. Zenodo, 2022. Li C, Li L, Geng Y, et al. Yolov6 v3. 0: A full-scale reloading[J]. arXiv preprint arXiv:2301.05586, 2023. Zhang H, Li D, Ji Y, et al. Toward new retail: A benchmark dataset for smart unmanned vending machines[J]. IEEE Transactions on Industrial informatics, 2019, 16(12): 7722-7731. Chen J, Wang S L, Lin H L. Out-of-stock detection based on deep learning[C]//Intelligent Computing Theories and Application: 15th International Conference, ICIC 2019, Nanchang, China, August 3–6, 2019, Proceedings, Part I 15. Springer International Publishing, 2019: 228-237. Arora D, Kulkarni K. Efficient Shelf Monitoring System using Faster-RCNN[C]//2024 International Conference on Distributed Computing and Optimization Techniques (ICDCOT). IEEE, 2024: 1-6. Zhao Y, Lv W, Xu S, et al. Detrs beat yolos on real-time object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 16965-16974. Liu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C]//Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part I 14. Springer International Publishing, 2016: 21-37. Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2980-2988. Lin T Y, Maire M, Belongie S, et al. Microsoft coco: Common objects in context[C]//Computer vision–ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13. Springer International Publishing, 2014: 740-755. Russell B C, Torralba A, Murphy K P, et al. LabelMe: a database and web-based tool for image annotation[J]. International journal of computer vision, 2008, 77: 157-173. Glenn Jocher et al. Yolov8: A comprehensive improvement of the yolo object detection series. https://docs.ultralytics.com/yolov8/, 2022. Wang C Y, Yeh I H, Mark Liao H Y. Yolov9: Learning what you want to learn using programmable gradient information[C]//European Conference on Computer Vision. Springer, Cham, 2025: 1-21. Wang A, Chen H, Liu L, et al. Yolov10: Real-time end-to-end object detection[J]. arXiv preprint arXiv:2405.14458, 2024. Jocher G, Qiu J. Ultralytics YOLO11. 2024[J]. URL https://github. com/ultralytics/ultralytics, 2024. Zhao Y, Lv W, Xu S, et al. Detrs beat yolos on real-time object detection[C]//Proceedings 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-6428418","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":449878960,"identity":"c085bb08-c26e-45c9-8af7-e8698c5bd5cc","order_by":0,"name":"Bidong Chen","email":"","orcid":"","institution":"Faculty of Applied Sciences, Macao Polytechnic University","correspondingAuthor":false,"prefix":"","firstName":"Bidong","middleName":"","lastName":"Chen","suffix":""},{"id":449878961,"identity":"10c9d0c0-ffa8-4e05-a605-741a661cb9f9","order_by":1,"name":"Lingui Li","email":"","orcid":"","institution":"Guangzhou College of Commerce","correspondingAuthor":false,"prefix":"","firstName":"Lingui","middleName":"","lastName":"Li","suffix":""},{"id":449878962,"identity":"96d5a2c9-a047-4138-a70f-4a1e09c0b124","order_by":2,"name":"Yuanda Lin","email":"","orcid":"","institution":"Whale TV Pte. (Singapore) Ltd","correspondingAuthor":false,"prefix":"","firstName":"Yuanda","middleName":"","lastName":"Lin","suffix":""},{"id":449878963,"identity":"137b3d37-3b38-4b5c-a304-c79a20b11b4f","order_by":3,"name":"Xu yang","email":"","orcid":"","institution":"Faculty of Applied Sciences, Macao Polytechnic University","correspondingAuthor":false,"prefix":"","firstName":"Xu","middleName":"","lastName":"yang","suffix":""},{"id":449878964,"identity":"1c286ecc-6f44-4c59-bac4-94deff79ceba","order_by":4,"name":"Sio Kei Im","email":"","orcid":"","institution":"Macao Polytechnic University","correspondingAuthor":false,"prefix":"","firstName":"Sio","middleName":"Kei","lastName":"Im","suffix":""},{"id":449878966,"identity":"a052269f-325c-4f18-be98-e11143580292","order_by":5,"name":"RuiPedro Paiva","email":"","orcid":"","institution":"University of Coimbra (CISUC), University of Coimbra","correspondingAuthor":false,"prefix":"","firstName":"RuiPedro","middleName":"","lastName":"Paiva","suffix":""},{"id":449878968,"identity":"4a25fc07-edac-4ea2-b7d7-19aa9bd786bc","order_by":6,"name":"Yapeng Wang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA80lEQVRIiWNgGAWjYHAD5gMQ+gDxWtgSSNbCY0CcFoPjZw8+Lvh1WN6cf83XzYVtDHJ8NxJYN3zAp+VMXrLxzL7DhjtnvN12e2Ybg7HkjQS2mzPwaTmQYybN23ObccONs9tu87YxJG4AarnNg0/L+Tfmv4Fa7DfcOPMMpKWesJYbOWbMPD9uJ24438MG0pJgQEiL5I13ydK8Df+TN9xgM7s945yE4cwzD9vw+oXvfO7Bzzx/0mw3nD/87HZBmY083/HkYzfwhZjCAaAbGNuALIkEBmYgCWQxNuDRwMAg3wBy9h8g5j8A0jIKRsEoGAWjABMAAGfxXUA/q+CSAAAAAElFTkSuQmCC","orcid":"","institution":"Faculty of Applied Sciences, Macao Polytechnic University","correspondingAuthor":true,"prefix":"","firstName":"Yapeng","middleName":"","lastName":"Wang","suffix":""}],"badges":[],"createdAt":"2025-04-11 12:38:30","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6428418/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6428418/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":81708811,"identity":"9c4f89fc-73a4-4550-8c45-3e7b311772f4","added_by":"auto","created_at":"2025-04-30 14:11:18","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":512922,"visible":true,"origin":"","legend":"\u003cp\u003eRPV11K: Products are placed at high density, and the vacancy features are not obvious; image acquisition angles are diversified\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6428418/v1/eac5fa16c2a3beb3be9ab997.png"},{"id":81708809,"identity":"a534e5ba-6c14-4215-8c8a-ebe93cf5590b","added_by":"auto","created_at":"2025-04-30 14:11:18","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":88958,"visible":true,"origin":"","legend":"\u003cp\u003eStatistical information of the RPV11K dataset\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6428418/v1/cb8cd6a7b38ec4de726171a7.png"},{"id":81708814,"identity":"5e1e0e75-55fb-42c7-8f98-6cf39b545a6e","added_by":"auto","created_at":"2025-04-30 14:11:19","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":228242,"visible":true,"origin":"","legend":"\u003cp\u003eRetail shelf product-vacancy detection pipeline: from data collection to edge deployment (DetectRPV)\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6428418/v1/7927fd47b65308cfb945df10.png"},{"id":81708812,"identity":"e5791f65-f6f8-483f-9024-bff9d4d9e869","added_by":"auto","created_at":"2025-04-30 14:11:18","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":53021,"visible":true,"origin":"","legend":"\u003cp\u003ePerformance comparison of various models on the RPV11K dataset\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6428418/v1/672485f5cd3fc8f2c746e8d7.png"},{"id":81708815,"identity":"9c0a9d1c-8577-4374-8c10-b6edf78e1387","added_by":"auto","created_at":"2025-04-30 14:11:19","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":715227,"visible":true,"origin":"","legend":"\u003cp\u003eQualitative detection results of the DetectRPV on the RPV11K dataset\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6428418/v1/aae4e3987a67f59fedf83430.png"},{"id":82642589,"identity":"299ff555-236d-4c25-ab81-aa0ea6f0f685","added_by":"auto","created_at":"2025-05-13 15:32:02","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2295062,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6428418/v1/3f3721df-b13b-4653-9b45-ecaaca41277b.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"RPV11K: A Benchmark for Joint Product-Vacancy Detection in Retail Scenarios","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eIn the era of digital transformation, intelligent retail is revolutionizing traditional business models through the integration of Artificial Intelligence (AI), robotics, and Internet of Things (IoT) technologies [1, 2]. This technological convergence is fundamentally reshaping shopping experiences and streamlining operational processes [3, 4]. Among\u0026nbsp;\u003c/p\u003e\n\u003cp\u003evarious retail operations, shelf management stands as a critical component that directly impacts business success through its operational efficiency and accuracy. Conventional shelf management methods depend predominantly on manual inspection, wherein personnel are required to physically assess aisles to ascertain product availability, location, and inventory levels [5, 6]. This labor-intensive approach is costly and subject to human error, resulting in challenges such as vacancy, misplaced products, and inefficient inventory management. The repercussions of merchandise shortages on customer satisfaction and brand reputation are substantial when consumers encounter such circumstances [7].\u003c/p\u003e\n\u003cp\u003eThe advent of unmanned retail establishments represents a paradigm shift in the domain of retail digitalization. These intelligent spaces employ sophisticated technology, including artificial intelligence (AI) powered computer vision systems, robotics, and the internet of things (IoT), to automate various operational aspects [4, 8]. Specifically, computer vision-based monitoring systems play a pivotal role in shelf management, vacancy detection, and automated replenishment. These systems possess the capability to initiate robotic replenishment upon detecting shelf vacancies, thereby establishing a completely automated workflow from the detection of vacancies to their resolution [8]. This automation substantially improves both operational efficiency and the overall shopping experience in retail environments.\u003c/p\u003e\n\u003cp\u003eHowever, the development of effective product and vacancy detection systems for real-world retail environments poses significant technical challenges. The unstructured nature of supermarket shelves, marked by dense product arrangements, diverse categories, and variable lighting conditions, gives rise to complex scenarios for detection algorithms. The dense arrangement of products gives rise to occlusion issues, while the diversity of products in terms of shape, size, and packaging increases the complexity of detection. Furthermore, varying illumination conditions, ranging from intense spotlights to dim aisles, have been shown to significantly affect the quality of visual information captured by monitoring systems. To address these challenges and advance the research into intelligent retail detection systems, our contributions are as follows:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eRPV11K Dataset: We introduce a shelf product and vacancy detection dataset containing 11,743 images and more than 1.87 million annotations, covering different retail scenarios, laying a solid foundation for future research.\u003c/li\u003e\n \u003cli\u003eExtensive Benchmarking: We conduct thorough evaluations using a series of lightweight object detection models at different scales, establishing baseline performance metrics for edge computing deployment. The benchmark provides valuable insights into model performance under real-world retail constraints.\u003c/li\u003e\n \u003cli\u003eEdge Computing Analysis: We explore model deployment on the Jetson-nano platform, providing preliminary performance analysis and practical observations for real-time retail detection systems. These findings may help bridge academic research and industrial applications.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe remainder of this paper is organized as follows: Section 2 reviews related work in object detection and retail applications. Section 3 details the proposed methodology. Section 4 presents experimental results and analysis. Finally, Section 5 concludes the paper with future directions.\u003c/p\u003e"},{"header":"2\tRelated work","content":"\u003ch2\u003e2.1 Intelligent retail\u003c/h2\u003e\n\u003cp\u003eIntelligent retail represents a significant advancement in the retail industry through the integration of advanced technologies to enhance customer experience and operational efficiency. Modern retail automation largely relies on artificial intelligence, IoT, and robotics technologies. Wang et al. [9] introduced an IoT-based inventory management system utilizing micro-sensors embedded in store shelves for product movement detection and inventory monitoring. While innovative, this approach faces practical limitations regarding deployment costs and maintenance requirements.\u003c/p\u003e\n\u003cp\u003eIn unmanned retail research, Zhang et al. [10] explored computer vision-based product recognition and pricing systems. Their approach employs deep learning algorithms for feature extraction and matching from product packaging. However, the system\u0026rsquo;s performance tends to decrease under challenging conditions, such as similar product packaging, partial occlusions, and varying lighting conditions.\u003c/p\u003e\n\u003cp\u003eWhile considerable research has addressed shopping process optimization, studies on automated shelf management, particularly vacancy detection, remain limited. Although Chen [11] and \u0026Scaron;ikić F [12] recognized the importance of shelf management, they have yet to propose effective solutions for product vacancy detection. This suggests potential opportunities for developing robust shelf management solutions.\u003c/p\u003e\n\u003ch2\u003e2.2 Existing dataset\u003c/h2\u003e\n\u003cp\u003eThe availability of data plays a foundational role in the advancement of intelligent retail technology, particularly through the utilization of open-source datasets. Early attempts to identify retail products had obvious flaws, particularly in terms of scale and scope. The LabelMe Food dataset [13], a pioneering achievement in image annotation research, contains a limited number of food product images with crowdsourced annotations, thereby restricting its applicability in retail scenarios. A similar limitation is exhibited by the Food101 dataset [14], which, despite its substantial size of 101,000 images across 101 categories, focuses exclusively on food items in controlled environments, thereby failing to capture the complexity of real retail settings.\u003c/p\u003e\n\u003cp\u003eAs the field develops, datasets of more complex scenarios gradually become available. The MS COCO-Stuff dataset [15], though not retail-specific, contributed valuable insights through its diverse object categories in various contexts. The Open Images Dataset [16] further expanded this scope, incorporating retail products among its numerous categories. Recent advancements have led to a shift towards real-world retail environments. The RP2K dataset [17] serves as a prime example. Although it is mainly centered around ecommerce applications, it offers valuable perspectives on the visual features of products in actual shopping scenarios. More recent contributions include the AiProducts-Challenge dataset [18,19], containing approximately 3 million images across 50,000 SKU-level categories, marking a significant advancement in scale. However, it presents challenges such as noisy data, unbalanced distribution, and annotation limitations. The SKU-110k dataset [20] specifically addresses dense product detection, while the RPC dataset [21] contributes 83,739 standardized images across 200 SKU products, captured through systematic photography methods.\u003c/p\u003e\n\u003cp\u003eOur research differs fundamentally by focusing on shelf vacancy detection, introducing unique challenges beyond traditional product recognition. The RPV11K dataset specifically addresses complex retail scenarios including product occlusion and dense placement, while uniquely incorporating both product and vacant area detection capabilities. This dual-category approach presents enhanced technical challenges compared to conventional single-category vacancy detection methods.\u003c/p\u003e\n\u003cp\u003eTo better demonstrate the uniqueness of our proposed RPV11K dataset, in addition to the datasets described above, we also make a comprehensive comparison with a variety of existing open-source datasets, as shown in Table 1. The comparison covers several key aspects, including dataset scale, annotation density, and detection capabilities. Our dataset shows significant advantages in terms of the number of images, bounding box density, and the integration of the most prominent product and vacancy detection features.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003cstrong\u003e.\u003c/strong\u003e Key attributes of relevant benchmarks\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"492\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 5px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 91px;\"\u003e\n \u003cp\u003eName\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eTime\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 69px;\"\u003e\n \u003cp\u003eImages\u003c/p\u003e\n \u003cp\u003equantity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 49px;\"\u003e\n \u003cp\u003eBbox\u003c/p\u003e\n \u003cp\u003equantity\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003eBox/image\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003eCategory\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 45px;\"\u003e\n \u003cp\u003eHigh\u003c/p\u003e\n \u003cp\u003eDensely\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003eVacancy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003eRetail\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 5px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 91px;\"\u003e\n \u003cp\u003ePASCAL VOC[39]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003e2012\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003e123287\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e896782\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e7.27\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e80\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 45px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 5px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 91px;\"\u003e\n \u003cp\u003eGrocery Shelves[4]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003e2015\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003e354\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e13184\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e37.24\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e10\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 45px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e✓\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 5px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 91px;\"\u003e\n \u003cp\u003eCOCO2017[40]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003e2017\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003e123287\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e52090\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e2.42\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e20\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 45px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"2\" style=\"width: 91px;\"\u003e\n \u003cp\u003eDota-v2.0[41]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003e2021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003e11268\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1793658\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e159.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 45px;\"\u003e\n \u003cp\u003e✓\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 5px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 91px;\"\u003e\n \u003cp\u003eSKU-110K[20]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"2\" valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003e2019\u003c/p\u003e\n \u003cp\u003e2021\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003e11762\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e1730996\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e147.17\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 45px;\"\u003e\n \u003cp\u003e✓\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e✓\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 5px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 91px;\"\u003e\n \u003cp\u003eLocount[42]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003e50394\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e809659\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e16.07\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e140\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 45px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e\u0026times;\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e✓\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 5px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 91px;\"\u003e\n \u003cp\u003ePRV12K(ours)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 37px;\"\u003e\n \u003cp\u003e2025\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 69px;\"\u003e\n \u003cp\u003e11743\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 49px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e1874077\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e159.59\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 57px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e2\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 45px;\"\u003e\n \u003cp\u003e✓\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 57px;\"\u003e\n \u003cp\u003e✓\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 37px;\"\u003e\n \u003cp\u003e✓\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eBbox quantity is short for bounding box, Box/image indicates the average number of bounding boxes per image; ✓ indicates that it has this attribute; \u003cstrong\u003e\u0026times;\u003c/strong\u003e indicates that there is no such attribute. The original SKU-110K has a portion of the image that is damaged and of poor quality.\u003c/p\u003e\n\u003ch2\u003e2.3 Retail Object Detection\u003c/h2\u003e\n\u003cp\u003eDeep learning-based methods have demonstrated significant advancements in product object detection. The YOLO series has established a prominent position in industrial applications, balancing real-time performance with detection accuracy. The initial YOLOv1, introduced by Redmon J et al. [22], innovatively approached object detection as a regression problem, enabling direct object localization and classification from global image features, thus substantially improving detection speed.\u003c/p\u003e\n\u003cp\u003eSubsequent YOLO iterations brought notable improvements. YOLOv5 [23] incorporated the CSPNet (Cross Stage Partial Network) structure to enhance feature propagation, achieving higher accuracy while maintaining detection efficiency. YOLOv6 [24] streamlined the network architecture through RepVGG structure, converting multi-branch features to single-branch structures during inference, thereby optimizing computational efficiency. YOLOv8 [34] further advanced performance through improved loss functions and data augmentation strategies, including enhanced CIoU loss for better multi-scale object detection.\u003c/p\u003e\n\u003cp\u003eDespite these advances, significant challenges persist in complex retail environments. In supermarket shelf scenarios characterized by densely arranged products and varying lighting conditions, these algorithms face performance limitations. Zhang et al. [25] identified that YOLO-based models often struggle with detection accuracy when encountering densely packed or partially occluded products.\u003c/p\u003e\n\u003ch2\u003e2.4 Vacancy Position Detection\u003c/h2\u003e\n\u003cp\u003eVacancy position detection represents a critical aspect of shelf management, yet research in this domain remains relatively nascent. Current approaches can be categorized into three main directions: image analysis, object detection algorithms, and multimodal data fusion. Image analysis approaches, exemplified by Chen et al. [26], employ background subtraction techniques to construct shelf background models and identify empty regions through comparative analysis. While innovative, these methods demonstrate significant sensitivity to lighting variations and often produce false detections in complex retail environments.\u003c/p\u003e\n\u003cp\u003eObject detection-based approaches offer an alternative methodology. Arora D et al. [27] implemented Faster R-CNN to detect shelf products and identify vacancies through layout comparison. However, real-world challenges including product occlusion, packaging similarities, and varying placement angles significantly impact detection reliability. While deep learning algorithms such as YOLO series, RT-DETR [28], SSD [29], and RetinaNet [30] have shown promising results in product void detection, they face persistent challenges in intelligent retail applications. These challenges primarily stem from the unstructured nature and environmental complexity of retail shelves, manifesting in:\u003c/p\u003e\n\u003cul\u003e\n \u003cli\u003eLimited detection accuracy in dense product arrangements;\u003c/li\u003e\n \u003cli\u003eInsufficient real-time performance in complex scenarios;\u003c/li\u003e\n \u003cli\u003eReduced reliability under varying lighting conditions;\u003c/li\u003e\n \u003cli\u003eDifficulty in handling partial occlusions.\u003c/li\u003e\n\u003c/ul\u003e\n\u003cp\u003eThe limitations of existing approaches underscore the necessity for more effective shelf vacancy detection methods. To address the challenges of complex retail environments and overcome the constraints of current solutions, this paper proposes an enhanced detection framework that integrates optimized object detection architecture with retail-specific features.\u003c/p\u003e"},{"header":"3\tMethod","content":"\u003ch2\u003e3.1 RPV11K dataset\u003c/h2\u003e\n\u003cp\u003eThe SKU-110K dataset [20] consists of 11,762 high-density retail scene images containing approximately 1.73 million annotated bounding boxes. The dataset is partitioned into 8,233 training images, 588 validation images, and 2,941 test images, all captured under standardized conditions with consistent device specifications and resolution. To enhance the dataset\u0026rsquo;s robustness and practical applicability, we implemented comprehensive data augmentation through random sampling techniques to address the limited lighting variations in the original dataset. The augmentation process was supplemented by incorporating additional images from complex retail scenarios, effectively expanding the environmental diversity of the training data. Furthermore, we refined the dataset composition to better align with current retail markets. This refinement involved removing product categories uncommon in chinese retail environments while substantially supplementing the dataset with prevalent shelf products. These modifications were specifically designed to enhance the dataset\u0026rsquo;s relevance to practical application scenarios in contemporary retail settings.\u003c/p\u003e\n\u003cp\u003eIn Fig.1, the RPV11K dataset systematically demonstrates the multi-dimensional characteristics of retail shelf environments through diverse perspective captures. Specifically, the dataset incorporates two critical viewing angles: the merchandise wall perspective, which provides a comprehensive macroscopic visualization of shelf layout configurations, and the front view perspective provides detailed visualization of product arrangement patterns on the shelf facade. This view captures a diverse array of products with distinctive chromatic characteristics. The side-angle perspective delineates the lateral structural composition of shelves and the hierarchical organization of product placement. The dual-perspective imaging upward and downward views enables comprehensive shelf analysis, capturing both detailed upper tier products and overall distribution patterns.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp; Additionally, the dataset incorporates random placement scenarios, authentically simulating the stochastic product arrangements that occur during routine retail operations such as inventory replenishment and promotional activities. This feature enhances both data authenticity and complexity. Regarding product density characteristics, the dataset shelves are populated with varied consumer goods categories including beverages, food items, and daily necessities. The high-density product arrangement accurately represents the characteristic spatial constraints of contemporary retail environments. In the context of vacancy detection, these diverse shelf scenarios and product placement variations provide comprehensive training materials. The multi-perspective shelf imagery and random placement situations effectively simulate real-world vacancy occurrences, facilitating the development and validation of detection algorithms. This robust data foundation ensures reliable technical support for practical applications in unmanned retail environments.\u003c/p\u003e\n\u003cp\u003eAs illustrated in Fig.2 (a), the quantitative relationship between product detection boxes and vacancy detection boxes exhibits a ratio of 12:1, indicating a predominant presence of product-related annotations relative to vacancy-related annotations within the dataset. Fig.2 (b) delineates the spatial distribution patterns of both product and vacancy detection boxes across individual images. Statistical analysis reveals that while products constitute the primary detection targets, the dataset encompasses partial vacancy position detection, demonstrating its comprehensive scope in addressing both product recognition and shelf vacancy monitoring. Heat map visualizations in Fig.2 (c) and Fig.2 (d) illustrate the spatial density distribution of product instances and shelf vacancies across the dataset. The spatial density analysis reveals a centrally concentrated distribution of product instances, with density gradually decreasing towards peripheral regions. In contrast, vacancy category detection boxes display distinct spatial distribution patterns, exhibiting an overall density degradation from central to peripheral regions, while maintaining localized high-density clusters.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003e3.2 Evaluation Metrics\u003c/h2\u003e\n\u003cp\u003eWe evaluate detection performance using standard metrics: Average Precision (AP) for single-class and mean Average Precision (mAP) for multi-class evaluation [20]. The metrics are defined as:\u003c/p\u003e\n\u003cp\u003e\u003cimg src=\"data:image/png;base64,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\" width=\"424\" height=\"237\"\u003e\u003c/p\u003e\n\u003cp\u003eWhere \u0026nbsp;TP , FP , FN \u0026nbsp;represent true positives, false positives, and false negatives respectively, and\u003cem\u003e\u0026nbsp;\u003c/em\u003e\u0026nbsp;N is the total number of classes.\u003c/p\u003e"},{"header":"4\tExperiments","content":"\u003ch2\u003e4.1 Experimental parameter settings\u003c/h2\u003e\n\u003cp\u003eWe employed different hyperparameter settings based on model scales. For nano and small-scale YOLO models, we set the weight decay parameter to 1e-2, with both initial and final learning rates maintained at 1e-3 and incorporated a Dropout ratio of 0.1. For medium-scale YOLO models, we increased the weight decay to 5e-2 while decreasing both initial and final learning rates to 1e-4 to ensure stable training convergence. Training was conducted on a server equipped with Intel(R) Xeon(R) Gold 6330 CPU and four NVIDIA A5000 GPUs. Data augmentation follows YOLOv11 guidelines to maintain scale invariance. The augmentation pipeline comprises Mosaic (4-image composition with scale range 0.5-1.5), random horizontal flipping (probability 0.5), and random erasure (10-30% of image area with aspect ratio 0.3-3.0). For optimization, nano-scale models (\u0026le;3M params) employ a higher weight decay (1e-2) to mitigate overfitting on small datasets, while medium models (20-60M params) adopt lower learning rates (1e-4) to stabilize gradient propagation through deeper networks.\u003c/p\u003e\n\u003cp\u003eAll model testing and performance evaluation were carried out exclusively on the Jetson Orin Nano Super platform (4GB RAM, 256GB storage) running Ubuntu 24.10. We adopte COCO-style metrics [31], including AP50 (average precision at IoU=50%), AP (averaged over IoU thresholds from 50% to 95% with a step of 5%), and mean Average Precision (mAP) for multi-class performance. Additionally, we utilized Precision, Recall, and F1-Score to assess detection capabilities and balance at different thresholds, while model operational efficiency was evaluated through Frames Per Second (FPS) and GFLOPs.\u003c/p\u003e\n\u003cp\u003eFig.3 outlines our DetectRPV framework. The annotation pipeline (Fig.3 a) begins with manual labeling using LabelMe [32], followed by format conversion to YOLO annotations for both product instances and shelf vacancies. The RPV11K dataset is partitioned into training, validation, and test sets following the established protocol of SKU-110K. We apply lossless compression techniques to reduce image storage requirements while preserving detection-critical features, optimizing inference speed without accuracy loss. We evaluate multiple YOLO variants on the processed dataset and deploy the optimized models on Jetson platform for real-time retail testing (Fig. 3 b).\u003c/p\u003e\n\u003ch2\u003e4.2 Results Analysis\u0026nbsp;\u003c/h2\u003e\n\u003cp\u003eDrawing upon the experimental results presented in Fig.4, we carried out a comprehensive performance assessment of various YOLO variants on the RPV11K dataset. In terms of latency-accuracy trade-off, YOLOv11 achieved the highest mAP value (approximately 55%) with a latency of 20ms, while YOLOv6 demonstrated balanced performance in terms of latency and accuracy. Regarding the comparison between model size and accuracy, YOLOv11 exhibited superior detection performance with a parameter configuration of 20.03M, validating the effectiveness of its architecture design. The training convergence curves showed that YOLOv11-m exhibited faster convergence rates and higher final accuracy compared to YOLOv9-s and YOLOv8-n. These experimental results provide important reference data for model selection in different application scenarios, and each model shows its applicability under different latency and accuracy requirements.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e2\u003c/strong\u003e\u003cstrong\u003e.\u003c/strong\u003e Performance comparison of different-scale models on Jetson edge computing platform, \u0026lsquo;-\u0026rsquo; denotes metrics unavailable due to edge device resource limitations.\u003c/p\u003e\n\u003cdiv align=\"\"\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"533\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eModels\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eScales\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 48px;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 48px;\"\u003e\n \u003cp\u003emAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 48px;\"\u003e\n \u003cp\u003emAP\u003csub\u003e50\u003c/sub\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eF1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 40px;\"\u003e\n \u003cp\u003elatency\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003eFPS\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 40px;\"\u003e\n \u003cp\u003eParams\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 48px;\"\u003e\n \u003cp\u003eGFLOPs\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv5[23]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003enano\u003c/p\u003e\n \u003cp\u003esmall\u003c/p\u003e\n \u003cp\u003emedium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.781\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.710\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.514\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.778\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.744\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e14.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e71\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e2.50\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e7.1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.792\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.730\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.534\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.794\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.760\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e24.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e40\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e7.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e18.7\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.794\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.743\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.546\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.804\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.768\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e48.5\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e22.13\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e52.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv6[24]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003enano\u003c/p\u003e\n \u003cp\u003eSmall\u003c/p\u003e\n \u003cp\u003emedium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.781\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.710\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.514\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.777\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.744\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e13.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e77\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e4.23\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e11.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.789\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.732\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.531\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.793\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.759\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e33.55\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e28\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e15.98\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e42.8\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.788\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.741\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.542\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; 0.799\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.764\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e51.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e158.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv8[33]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003enano\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003esmall\u003c/p\u003e\n \u003cp\u003emedium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.787\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.714\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.520\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.782\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.749\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e15.6\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e64\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e3.01\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e8.1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.785\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.719\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.519\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.783\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.751\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e28.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e35\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e9.82\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e23.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.791\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.744\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.546\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.803\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.767\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e54.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e18.25\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e23.22\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e67.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv9[34]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003etiny\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003esmall\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003emedium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.785\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.712\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.520\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.782\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.747\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e23.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e43\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e1.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e6.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.792\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.738\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.542\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.800\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.764\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e35.7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e28\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e6.19\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e22.1\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.789\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.737\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.536\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.796\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.762\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e57.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e20.01\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e76.5\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv10[35]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003enano\u003c/p\u003e\n \u003cp\u003esmall\u003c/p\u003e\n \u003cp\u003emedium\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.781\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.705\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.514\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.776\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.741\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e17.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e56\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e2.70\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e8.2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.787\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.724\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.533\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.790\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.754\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e31\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e32\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e8.03\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e24.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.790\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.737\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.544\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.798\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.763\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e54\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e19\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e16.48\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e63.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"3\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv11[36]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd rowspan=\"3\" valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003enano\u003c/p\u003e\n \u003cp\u003esmall\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003emedium\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.782\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.713\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.516\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.779\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.746\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e17.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e2.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e6.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.793\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.730\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.534\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.795\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.760\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e30.0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e33\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e9.41\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e21.3\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.797\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.744\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.551\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.806\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e0.770\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e52.0\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e19\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e20.03\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e67.7\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 72px;\"\u003e\n \u003cp\u003eRT-DETR[37]\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.763\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.689\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.487\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e0.737\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.724\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e\u003cstrong\u003e-\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 32px;\"\u003e\n \u003cp\u003e-\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 40px;\"\u003e\n \u003cp\u003e63.18\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 48px;\"\u003e\n \u003cp\u003e103.4\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n\u003c/div\u003e\n\u003cp\u003eTable 2 presents the comparative evaluation of YOLO variants and RT-DETR on RPV11K. YOLOv11-medium achieves superior performance across all metrics, with precision of 0.797, recall of 0.744, F1-score of 0.770, and mAP50 of 0.806. YOLOv8-medium and YOLOv9-medium show competitive results, while RT-DETR-large yields relatively lower precision (0.763). Notably, lightweight models like YOLOv9-tiny and YOLOv8-nano demonstrate strong recall performance (0.712 and 0.714 respectively). Table 3 summarizes the benchmark experiments presented in Table 2, covering the performance of models of different scales. In product detection, YOLOv9-s and YOLOv11-m demonstrate outstanding performance with remarkable precision, recall, and F1-scores, while YOLOv8-n is slightly less remarkable. Regarding product vacancy detection, the precision and recall of all models are relatively low, indicating that there is substantial room for improvement, which may be attributed to the challenging nature of vacancy features. In terms of average precision, YOLOv11-m performs best in product detection. Overall, model optimization is necessary to enhance the vacancy detection ability and contribute to the optimization of retail management.\u003c/p\u003e\n\u003cp\u003eIn terms of model complexity, the RT-DETR model has the largest parameter scale and the highest computational load, while the YOLOv5 nano model has the smallest parameter scale and computational load. The YOLOv11 medium model achieves a good balance between performance and complexity, showing great potential. Other models also have their own advantages in different metrics. Thus, they can be selected according to specific requirements. In the future, exploring the optimization of the model structure can be considered to improve performance while reducing complexity.\u003c/p\u003e\n\u003cp\u003eFig.5 shows the results of qualitative testing of the RPV11K dataset through the DetectRPV process. We observed a series of supermarket shelf images, which were filled with various commodities, including food, daily necessities, etc. The commodities were rich and varied, colorful, and neatly arranged. Through the DetectRPV test process, the location of the commodities in the image was clearly marked, and each commodity was circled with a different color box, which intuitively presented the model\u0026apos;s detection effect on the commodity. From the test results, the YOLOv11 benchmark model can accurately identify commodities and detect vacant locations in complex scenes such as supermarket shelves, but there is still a lot of room for improvement. It is worthwhile for us to continue to use RPV11K as a benchmark to study models with better performance.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003e3\u003c/strong\u003e\u003cstrong\u003e.\u003c/strong\u003e Performance benchmark models of different sizes on RPV11K data\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" align=\"left\" width=\"393\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 72px;\"\u003e\n \u003cp\u003eModels\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eClass\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 60px;\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003emAP\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003emAP\u003csub\u003e50\u003c/sub\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eF1\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv8-n\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eProduct\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 60px;\"\u003e\n \u003cp\u003e0.890\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.863\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.593\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.917\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.876\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eVacancy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 60px;\"\u003e\n \u003cp\u003e0.685\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.564\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.447\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.619\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv9-s\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eProduct\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 60px;\"\u003e\n \u003cp\u003e0.897\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.886\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.613\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.932\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.891\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eVacancy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 60px;\"\u003e\n \u003cp\u003e0.686\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.589\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.470\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.669\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.636\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd rowspan=\"2\" style=\"width: 72px;\"\u003e\n \u003cp\u003eYOLOv11-m\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eProduct\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 60px;\"\u003e\n \u003cp\u003e0.897\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.900\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.622\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.938\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.898\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 52px;\"\u003e\n \u003cp\u003eVacancy\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 60px;\"\u003e\n \u003cp\u003e0.698\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.589\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.480\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.674\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 52px;\"\u003e\n \u003cp\u003e0.639\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"5\tConclusion","content":"\u003cp\u003eCurrent object detectors, while successful in standard benchmarks, face significant challenges in dense retail environments. To address this gap, we introduce RPV11K - a new benchmark dataset of retail shelf images with precise annotations for both product detection and vacancy locations. Our work makes two key contributions: (1) establishing a challenging retail-focused dataset that reflects real-world complexity, and (2) conducting comprehensive evaluations of lightweight detectors suitable for edge deployment in shelf management robots. Our experiments on the Jetson-nano platform demonstrate that even state-of-the-art models like YOLOv11 achieve a limited performance, indicating substantial room for improvement in dense retail scene understanding. The RPV11K dataset, as the first large-scale benchmark focusing on retail product and shelf vacancy detection, provides a reliable evaluation framework for developing and accessing automated retail solutions, facilitating technological advancement in this domain.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e.\u003cstrong\u003e\u0026nbsp;\u003c/strong\u003eThis work was supported by Foundation for Science and Technology (FCT) through national funds (UIDB/00326/2025, UIDP/00326/2025) and Macao Polytechnic University (Grant No. RP/FCA-04/2022).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of Competing Interest.\u003c/strong\u003e All authors declare no conflicts of interest; All authors read and approved the final manuscript; No historical conflicts of interest; No conflicts of interest in the past 3 years. \u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData availability.\u003c/strong\u003e The data is constructed by ourselves, and there has been no conflict of interest in history. The experimental data in the paper will be provided upon request.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIf necessary, you can contact corresponding author:\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e(Correspondence: L.L, [email protected]; Y.W, [email protected];)\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eCredit authorship contribution statement. All authors read, were informed of, and approved the final manuscript. Contributions are as follows: Conceptualization,B.C,L.L,Y.L;\u003c/p\u003e\n\u003cp\u003emethodology,B.C,L.L; software,B.C,L.L validation,B.C,\u003c/p\u003e\n\u003cp\u003eformal analysis,B.C,Y.W; investigation,B.C,L.L,Y.L;\u003c/p\u003e\n\u003cp\u003eresources,B.C,L.L data curation,B.C,L.L; writing\u0026mdash;original draft preparation,B.C;\u003c/p\u003e\n\u003cp\u003ewriting\u0026mdash;review and editing,B.C,L.L; visualization,B.C,L.L;\u003c/p\u003e\n\u003cp\u003esupervision,X.Y,S.I,R.P,Y.W; project administration,B.C,X.Y,S.I,R.P,Y.W;\u003c/p\u003e\n\u003cp\u003efunding acquisition,X.Y,R.P,Y.W; \u0026nbsp;All authors have read and agreed to the published version of the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDisclosure of Interests.\u003c/strong\u003e The authors declare no conflict of interest.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003eGauri D K, Jindal R P, Ratchford B, et al. Evolution of retail formats: Past, present, and future[J]. Journal of Retailing, 2021, 97(1): 42-61.\u003c/li\u003e\n \u003cli\u003eHar L L, Rashid U K, Te Chuan L, et al. Revolution of retail industry:\u0026nbsp;from perspective of retail 1.0 to 4.0[J]. Procedia Computer Science,\u0026nbsp;2022, 200: 1615-1625.\u003c/li\u003e\n \u003cli\u003eRane N. Integrating leading-edge artificial intelligence (AI), internet\u0026nbsp;of things (IOT), and big data technologies for smart and sustainable\u0026nbsp;architecture, engineering and construction (AEC) industry: Challenges\u0026nbsp;and future directions[J]. Engineering and Construction (AEC) Industry:\u0026nbsp;Challenges and Future Directions (September 24, 2023), 2023.\u003c/li\u003e\n \u003cli\u003eWolniak R, Stecuła K, Aydın B. Digital Transformation of Grocery In-Store Shopping-Scanners, Artificial Intelligence, Augmented Reality and\u0026nbsp;Beyond: A Review[J]. Foods, 2024, 13(18): 2948.\u003c/li\u003e\n \u003cli\u003eH\u0026uml;ubner, A. (2011). Retail Category Management: Decision Support\u0026nbsp;Systems for Assortment, Shelf Space, Inventory and Price Planning.\u0026nbsp;Springer Science \u0026amp; Business Media.\u003c/li\u003e\n \u003cli\u003eSajadi S J, Ahmadi A. An integrated optimization model and meta-heuristics for assortment planning, shelf space allocation, and inventory\u0026nbsp;management of perishable products: a real application[J]. Plos one, 2022,17(3): e0264186.\u003c/li\u003e\n \u003cli\u003eKollock P. The production of trust in online markets[J]. Advances in\u0026nbsp;group processes, 1999, 16(1): 99-123.\u003c/li\u003e\n \u003cli\u003eVermesan O, Br\u0026uml;oring A, Tragos E, et al. Internet of robotic\u0026nbsp;things\u0026ndash;converging sensing/actuating, hyperconnectivity, artificial intelligence and IoT platforms[M]//Cognitive hyperconnected digital transformation. River Publishers, 2022: 97-155.\u003c/li\u003e\n \u003cli\u003eWang Y C, Yang C C. 3S-cart: a lightweight, interactive sensor-based\u0026nbsp;cart for smart shopping in supermarkets[J]. IEEE Sensors Journal, 2016,16(17): 6774-6781.\u003c/li\u003e\n \u003cli\u003eZhang H, Li D, Ji Y, et al. Toward new retail: A benchmark dataset for\u0026nbsp;smart unmanned vending machines[J]. IEEE Transactions on Industrial\u0026nbsp;informatics, 2019, 16(12): 7722-7731.\u003c/li\u003e\n \u003cli\u003eChen J, Wang S L, Lin H L. Out-of-stock detection based on deep\u0026nbsp;learning[C]//Intelligent Computing Theories and Application: 15th In-ternational Conference, ICIC 2019, Nanchang, China, August 3\u0026ndash;6, 2019,\u0026nbsp;Proceedings, Part I 15. Springer International Publishing, 2019: 228-237.\u003c/li\u003e\n \u003cli\u003eˇSiki\u0026acute;c F, Kalafati\u0026acute;c Z, Subaˇsi\u0026acute;c M, et al. Enhanced Out-of-Stock Detection\u0026nbsp;in Retail Shelf Images Based on Deep Learning[J]. Sensors, 2024, 24(2):693.\u003c/li\u003e\n \u003cli\u003eRussell B C, Torralba A, Murphy K P, et al. LabelMe: a database\u0026nbsp;and web-based tool for image annotation[J]. International journal of\u0026nbsp;computer vision, 2008, 77: 157-173.\u003c/li\u003e\n \u003cli\u003eBossard L, Guillaumin M, Van Gool L. Food-101\u0026ndash;mining discriminative\u0026nbsp;components with random forests[C]//Computer vision\u0026ndash;ECCV 2014:13th European conference, zurich, Switzerland, September 6-12, 2014,\u0026nbsp;proceedings, part VI 13. Springer International Publishing, 2014: 446-461.\u003c/li\u003e\n \u003cli\u003eCaesar H, Uijlings J, Ferrari V. Coco-stuff: Thing and stuff classes in context[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2018: 1209-1218.\u003c/li\u003e\n \u003cli\u003eKrasin I, Duerig T, Alldrin N, et al. Openimages: A public dataset for large-scale multi-label and multi-class image classification[J]. Dataset available from https://github. com/openimages, 2017, 2(3): 18.\u003c/li\u003e\n \u003cli\u003ePeng J, Xiao C, Li Y. RP2K: A large-scale retail product dataset for fine-grained image classification[J]. arXiv preprint arXiv:2006.12634, 2020.\u003c/li\u003e\n \u003cli\u003eLiang Y, Zhang W. Winner Solution for AliProducts Challenge: Large-scale Product Recognition[J].\u003c/li\u003e\n \u003cli\u003eCheng L, Zhou X, Zhao L, et al. Weakly supervised learning with side information for noisy labeled images[C]//Computer Vision\u0026ndash;ECCV 2020: 16th European Conference, Glasgow, UK, August 23\u0026ndash;28, 2020, Proceedings, Part XXX 16. Springer International Publishing, 2020: 306-321.\u003c/li\u003e\n \u003cli\u003eE. Goldman, R. Herzig, A. Nachman, M. Salzmann, and D. Freedman. Precise detection in densely packed scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5227-5236, 2019.\u003c/li\u003e\n \u003cli\u003eWei X S, Cui Q, Yang L, et al. RPC: A large-scale retail product checkout dataset[J]. arXiv preprint arXiv:1901.07249, 2019.\u003c/li\u003e\n \u003cli\u003eRedmon J. You only look once: Unified, real-time object detection[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2016.\u003c/li\u003e\n \u003cli\u003eJocher G, Chaurasia A, Stoken A, et al. ultralytics/yolov5: v7. 0-yolov5 sota realtime instance segmentation[J]. Zenodo, 2022.\u003c/li\u003e\n \u003cli\u003eLi C, Li L, Geng Y, et al. Yolov6 v3. 0: A full-scale reloading[J]. arXiv preprint arXiv:2301.05586, 2023.\u003c/li\u003e\n \u003cli\u003eZhang H, Li D, Ji Y, et al. Toward new retail: A benchmark dataset for smart unmanned vending machines[J]. IEEE Transactions on Industrial informatics, 2019, 16(12): 7722-7731.\u003c/li\u003e\n \u003cli\u003eChen J, Wang S L, Lin H L. Out-of-stock detection based on deep learning[C]//Intelligent Computing Theories and Application: 15th International Conference, ICIC 2019, Nanchang, China, August 3\u0026ndash;6, 2019, Proceedings, Part I 15. Springer International Publishing, 2019: 228-237.\u003c/li\u003e\n \u003cli\u003eArora D, Kulkarni K. Efficient Shelf Monitoring System using Faster-RCNN[C]//2024 International Conference on Distributed Computing and Optimization Techniques (ICDCOT). IEEE, 2024: 1-6.\u003c/li\u003e\n \u003cli\u003eZhao Y, Lv W, Xu S, et al. Detrs beat yolos on real-time object detection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 16965-16974.\u003c/li\u003e\n \u003cli\u003eLiu W, Anguelov D, Erhan D, et al. Ssd: Single shot multibox detector[C]//Computer Vision\u0026ndash;ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11\u0026ndash;14, 2016, Proceedings, Part I 14. Springer International Publishing, 2016: 21-37.\u003c/li\u003e\n \u003cli\u003eLin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[C]//Proceedings of the IEEE international conference on computer vision. 2017: 2980-2988.\u003c/li\u003e\n \u003cli\u003eLin T Y, Maire M, Belongie S, et al. Microsoft coco: Common objects in context[C]//Computer vision\u0026ndash;ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part v 13. Springer International Publishing, 2014: 740-755.\u003c/li\u003e\n \u003cli\u003eRussell B C, Torralba A, Murphy K P, et al. LabelMe: a database and web-based tool for image annotation[J]. International journal of computer vision, 2008, 77: 157-173.\u003c/li\u003e\n \u003cli\u003eGlenn Jocher et al. Yolov8: A comprehensive improvement of the yolo object detection series. https://docs.ultralytics.com/yolov8/, 2022.\u003c/li\u003e\n \u003cli\u003eWang C Y, Yeh I H, Mark Liao H Y. Yolov9: Learning what you want to learn using programmable gradient information[C]//European Conference on Computer Vision. Springer, Cham, 2025: 1-21.\u003c/li\u003e\n \u003cli\u003eWang A, Chen H, Liu L, et al. Yolov10: Real-time end-to-end object detection[J]. arXiv preprint arXiv:2405.14458, 2024.\u003c/li\u003e\n \u003cli\u003eJocher G, Qiu J. Ultralytics YOLO11. 2024[J]. URL https://github. com/ultralytics/ultralytics, 2024.\u003c/li\u003e\n \u003cli\u003eZhao Y, Lv W, Xu S, et al. Detrs beat yolos on real-time object detection[C]//Proceedings \u003c/li\u003e\n\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":"Intelligent retail, vacancy detection, product detection, edge deployment, shelf monitoring","lastPublishedDoi":"10.21203/rs.3.rs-6428418/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6428418/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"In intelligent retail, accurate detection of densely arranged products and shelf vacancies in unstructured environments remains a critical challenge. This paper introduces RPV11K, a large-scale benchmark dataset (11,743 im-ages, 1.87M annotations) designed for joint product and vacancy detection in real-world retail scenarios. We develop a novel edge-deployable framework, DetectRPV, integrating lightweight YOLO variants and systematic data augmentation. Experimental results on the Jetson edge platform show that YOLOv1-medium achieves the best performance with mAP50 of 80.6% and mAP of 55.1%, while maintaining an inference speed of 52ms. RPV11K, the first dataset to explicitly model both product and vacancy categories under dense occlusion and diverse lighting, provides a rigorous evaluation benchmark for automated shelf management systems. Our work bridges the gap between academic research and industrial deployment by establishing a benchmark for real-time shelf monitor-ing with metrics relevant to automated retail operations. The code and data will be made available on https://github.com/Cy1oong/RPV11K.","manuscriptTitle":"RPV11K: A Benchmark for Joint Product-Vacancy Detection in Retail Scenarios","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-30 14:11:14","doi":"10.21203/rs.3.rs-6428418/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":"06eb6848-2c1e-4e66-bd0e-2b729ff82b90","owner":[],"postedDate":"April 30th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":47874159,"name":"Physical sciences/Engineering/Electrical and electronic engineering"},{"id":47874160,"name":"Physical sciences/Mathematics and computing"},{"id":47874161,"name":"Physical sciences/Mathematics and computing/Applied mathematics"},{"id":47874162,"name":"Physical sciences/Mathematics and computing/Computational science"},{"id":47874163,"name":"Physical sciences/Mathematics and computing/Computer science"},{"id":47874164,"name":"Physical sciences/Mathematics and computing/Information technology"},{"id":47874165,"name":"Physical sciences/Mathematics and computing/Pure mathematics"},{"id":47874166,"name":"Physical sciences/Mathematics and computing/Scientific data"},{"id":47874167,"name":"Physical sciences/Mathematics and computing/Software"},{"id":47874168,"name":"Physical sciences/Mathematics and computing/Statistics"}],"tags":[],"updatedAt":"2025-05-13T15:23:53+00:00","versionOfRecord":[],"versionCreatedAt":"2025-04-30 14:11:14","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6428418","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6428418","identity":"rs-6428418","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}