Novel Mean Track Length-Driven 3D Framework via COLMAP for Architectural Heritage Photogrammetry

Novel Mean Track Length-Driven 3D Framework via COLMAP for Architectural Heritage Photogrammetry | 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 Novel Mean Track Length-Driven 3D Framework via COLMAP for Architectural Heritage Photogrammetry Luo Wang, Yating Duan, Jungang Jiang, Yiheng Liu, Hailin Zhang, and 2 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8709145/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Existing photogrammetric workflows for architectural heritage documentation often lack process-oriented quantitative criteria to guide image acquisition, resulting in reconstruction failure due to data sparsity or noise and inefficiency caused by excessive image acquisition. This study proposes a quantitative threshold framework based on Mean Track Length (MTL) for component-level 3D reconstruction using COLMAP. A total of 66 reconstruction experiments and 462 data records were conducted on eight types of traditional architectural components. The results indicate that reliable reconstruction is achieved when MTL lies within 3.60–4.20; values below this range lead to unstable geometry, whereas values above it do not improve reconstruction quality and instead introduce noise. This study establishes a measurable stopping criterion for image acquisition and transforms reconstruction from an experience-driven practice into a threshold-driven and controllable workflow, enabling reproducible and efficient architectural heritage documentation. Architectural heritage Photogrammetry Mean Track Length Quantification threshold framework COLMAP Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 1. Introduction 1.1. Research background As pressures on cultural heritage conservation continue to increase, digital technologies—particularly three-dimensional (3D) modelling, point cloud generation, and virtual visualization—have played an increasingly important role in heritage documentation, conservation, and management[1]. Previous studies have shown that 3D digital technologies can generate highly detailed digital heritage models and provide essential baseline data to support long-term preservation, restoration planning, and public education[2]. Moreover, digital heritage practices are progressively extending toward virtual environments and public engagement, enhancing the global accessibility and interpretability of cultural heritage through digital platforms[3]. Empirical studies conducted in China likewise highlight the significance of digital technologies in the conservation of heritage buildings, emphasizing their advantages in condition recording, decision support, and the development of interactive heritage experiences[4]. With the continuous advancement of cultural heritage conservation and digital technologies, the digital documentation of architectural heritage has become an important research focus in the fields of architecture, archaeology, and heritage conservation[5]. Digital technologies—particularly image-based three-dimensional (3D) modelling—enable non-contact and accurate recording of the geometric forms, surface details, and texture characteristics of heritage assets, providing critical support for long-term preservation, restoration planning, and virtual presentation[6]. Through digital methods, architectural heritage can be virtually reconstructed and systematically archived, while precise 3D data can be generated to support restoration interventions, effectively reducing the risk of damage associated with physical contact[7]. 1.2. Literature review Existing approaches to three-dimensional documentation of architectural heritage can be broadly categorized into three main pathways. Three-dimensional laser scanning enables rapid acquisition of high-density, high-precision point cloud data and offers clear advantages in capturing the overall geometry of large-scale buildings and complex spatial environments. However, the high cost of equipment, along with demanding operational and post-processing requirements create significant barriers to entry, thereby stifling its widespread adoption in routine heritage surveys[8]. Image-based photogrammetric techniques reconstruct three-dimensional models from multi-view imagery without relying on dedicated surveying equipment, offering a comparatively low-cost solution. As a result, photogrammetry has become one of the most widely adopted approaches in digital heritage research. Nevertheless, many mature software platforms and algorithmic frameworks are characterized by steep learning curves or high commercial licensing costs, which continue to limit their accessibility and broader application[9]. Unmanned aerial vehicle (UAV)-based image acquisition and oblique photogrammetry are well suited for the overall documentation of large-scale heritage assets. However, these approaches are primarily restricted to outdoor environments and impose additional requirements related to flight control and airspace regulation, which continue to limit their applicability in certain heritage documentation contexts[10]. Although structure-from-motion (SfM) techniques offer clear advantages in terms of cost efficiency and equipment requirements, existing studies have predominantly focused on post-reconstruction model accuracy assessment or algorithmic improvements. This reliance on post-hoc evaluation offers limited decision support during the acquisition phase, as it lacks real-time, process-oriented quantitative metrics essential for preempting reconstruction failure. Crucially, conventional metrics like reprojection error often remain low even in geometrically unstable models due to solver optimization, masking underlying data sufficiency issues. Often, by the time quality issues are detected in post-processing, the field team has already left the heritage site, making re-acquisition impossible. In practical applications, this deficiency often leads to reconstruction failures caused by insufficient image acquisition, or, conversely, to excessive data redundancy and increased computational costs due to over-acquisition. Therefore, there is a pressing need to introduce quantitative indicators that provide quality feedback during the reconstruction process, enabling dynamic regulation of image acquisition and early prediction of modelling reliability[11]. In addition, the integration of multiple techniques—such as laser scanning, photogrammetry, and UAV-based acquisition—has become an emerging trend in comprehensive digital heritage documentation. However, such integrated approaches are typically associated with higher overall costs and increased workflow complexity[12]. Although numerous studies have demonstrated the respective advantages of these technologies in heritage reconstruction, most existing research has focused primarily on macro-scale architectural structures, overall scenes, or large heritage sites. At the component and detail levels, there is still a lack of targeted threshold definitions, standardized workflows, and low-cost, scalable solutions for high-precision modelling[13]. Moreover, substantial barriers related to resources, equipment, and professional expertise remain, further constraining the widespread adoption and practical dissemination of these techniques. 1.3. Research recaps and gaps Compared with other open-source photogrammetric tools, COLMAP demonstrates notable advantages in the robustness of feature extraction and matching, the modelling of multi-view geometric constraints, and the accessibility of process-oriented metrics. These characteristics enable COLMAP not only to support 3D model generation, but also to facilitate process-oriented analysis of reconstruction stability, making it particularly suitable for component-level image-based 3D reconstruction research[14]. Building upon this, this study establishes a low-cost, open-source modelling workflow centered on COLMAP and MeshLab[15, 16]. Instead of pursuing the absolute highest geometric accuracy, this workflow prioritizes the cost-effectiveness ratio and process reliability for rapid heritage documentation. Based on extensive comparative experiments, and considering variations in component size, texture strength, geometric structure, and color characteristics, a set of practical modelling threshold ranges is proposed to assist in balancing image redundancy, reconstruction stability, and modelling efficiency, thereby avoiding redundancy effects and reducing unnecessary data acquisition and computational costs[17]. In addition, a frame extraction strategy is introduced to simplify the data acquisition process, making image capture and modelling workflows more controllable and providing a threshold-driven, risk-aware decision pathway for component-level image-based modelling. This study aims to promote the application of lightweight, low-cost, and reproducible digital documentation methods for architectural heritage in grassroots surveys and educational contexts, offering quantitative references for determining when image-based modelling becomes sufficiently reliable at the component level, and supporting the standardization and methodological development of heritage digitization technologies. 2. Methodology 2.1. Research workflow This study introduces a low-cost, easily reproducible, and parameter-fixed open-source workflow designed to provide a practical and affordable digital modelling solution for traditional architectural components. To clearly illustrate the proposed workflow, the overall digital modelling process is summarized in Fig. 1 . This figure presents the complete pipeline, including image acquisition and preprocessing, feature extraction, matching, and three-dimensional reconstruction, followed by mesh generation, texture mapping, and quality evaluation. 2.2. Data acquisition and preprocessing Image acquisition was performed using standard consumer-grade smartphones to capture photographs of architectural components. To ensure data completeness and operational efficiency, a multi-view acquisition strategy was employed[18], whereby components were densely imaged from multiple viewpoints along surrounding trajectories. This method guaranteed comprehensive coverage of each component while capturing sufficient texture information for reliable three-dimensional reconstruction. For components with complex textures and fine-scale details, particular attention was given to minimizing the impact of illumination variations on the reconstruction results. All components were photographed using the same smartphone to maintain consistency in image properties, including resolution, focal length, and aperture. Additionally, care was taken to preserve the original image metadata during data transfer and format conversion, thereby preventing the loss of critical information such as pixel resolution, focal length, and aperture settings. To ensure reproducibility in subset generation, a standardized sequential renaming protocol was applied to all valid images prior to ingestion into the reconstruction pipeline. After the renaming procedure, a uniform frame extraction strategy was applied to the high-density image sets. Initially, images were subsampled by selecting files with sequential indices of 1, 3, 5, and so on, to form the first dataset. Additional datasets were then generated by extracting images with indices of 1, 4, 7, 11, etc., thereby producing multiple image subsets with varying quantities of images. The implementation code for the frame-sampling method is provided in the Supplementary Materials. By varying the sampling step size S, multiple image subsets containing different numbers of images were generated, allowing for analysis of the impact of image quantity on three-dimensional reconstruction quality. This frame extraction strategy facilitates the creation of multiple image datasets of varying sizes under consistent acquisition conditions, eliminating the need for repeated image collection. The validation results demonstrating the effectiveness of the frame extraction strategy are presented in Table 1 . This frame extraction strategy facilitates the creation of multiple image datasets of varying sizes without the need for repeated physical collection. Crucially, this approach allows for a controlled simulation of varying acquisition densities under identical illumination and sensor conditions, thereby isolating image quantity as the sole variable affecting reconstruction quality. Table 1 Validation of the effectiveness of the frame extraction strategy. Note: “Saturated” is defined as the state where additional images increase file size and processing time without yielding any perceptible improvement in the Visual Quality Score (maintaining a score of 5.0). Component ID Frame extraction strategy Number of images Mean Track Length Reconstruction quality Within effective range Component 3 Full acquisition (no frame extraction) 288 5.40 Saturated No (redundant) Component 3 S = 2 144 4.64 Saturated No (redundant) Component 3 S = 3 96 4.25 Acceptable Yes Component 3 S = 4 72 4.10 Acceptable Yes Component 3 S = 5 58 3.86 Acceptable Yes Component 3 S = 6 48 3.63 Acceptable Yes Component 3 S = 7 42 3.58 Unacceptable No 2.3. Key parameter settings During feature extraction, matching, and three-dimensional reconstruction with COLMAP, default parameter settings were consistently used to enhance the reproducibility of the proposed workflow. However, in practical applications—especially for low-texture components or image sets with limited overlap, where matching and reconstruction are more challenging—certain parameters can be adjusted to improve feature acquisition and reconstruction robustness. The following parameter adjustments serve as empirical heuristics for challenging scenarios and act as a supplement to the core threshold-driven workflow. During the feature extraction stage, the parameter max_image_size determines the target resolution to which images are uniformly rescaled before feature extraction. At lower resolutions, fewer feature points are typically detected, while increasing this value enables the extraction of more features but also raises computational demands. Similarly, max_num_features specifies the maximum number of feature points that can be extracted from each image; increasing this parameter can improve matching robustness by providing more features, but it also results in greater memory usage and longer processing times. Therefore, both parameters should be adjusted based on the characteristics of the image set and the available computational resources. During the feature matching stage, the parameter max_num_matches sets the maximum number of feature correspondences established between each image pair. Increasing this value allows more matches to be retained, which may enhance matching completeness but also leads to greater memory consumption. Therefore, max_num_matches should be adjusted according to the specific characteristics of the image set and the available computational resources. During mesh generation in MeshLab, considering both modelling quality and hardware constraints, only the Reconstruction Depth parameter was adjusted to a value of 15, while all other parameters remained at their default settings. 2.4. Data storage After completing dense point cloud reconstruction and subsequent cleaning, the point cloud data exported from COLMAP were imported into MeshLab for mesh reconstruction and texture mapping. First, a manual filtering process was conducted to remove background points unrelated to the architectural component, retaining only those corresponding to the target component. This step helps prevent redundant data from interfering with subsequent mesh generation and texture mapping procedures. After point cloud cleaning, mesh reconstruction was performed. In MeshLab, a continuous triangular mesh surface was generated from the oriented point cloud using the menu path Filters → Remeshing, Simplification and Reconstruction → Surface Reconstruction: Screened Poisson. This method effectively suppresses the influence of point cloud noise on the mesh structure while preserving overall geometric continuity, thus providing a stable geometric foundation for subsequent texture mapping. Prior to texture parameterization and mapping, representative original images intended for texture mapping were imported into MeshLab. To ensure high-quality texture mapping, the image dataset was selected to cover key component regions and capture typical texture characteristics. The selected images were loaded into the system via File → Import Raster, providing the data source for subsequent image-based texture mapping procedures. After mesh generation and image data import, texture parameterization and mapping were performed. First, under Filters → Texture, the “Parameterization + texturing from registered rasters” function was used to complete texture mapping with the registered multi-view original images. The image information was projected onto the mesh surface, resulting in a textured model with a realistic visual appearance. After completing mesh reconstruction and texture mapping, the model can be exported based on subsequent application requirements. MeshLab supports various common 3D model formats, such as .3ds, .obj, and .glb, offering strong compatibility and flexibility for further processing, visualization, and cross-platform applications. 2.5. Evaluation and result validation To systematically evaluate the effectiveness, stability, and scalability of the proposed lightweight modelling framework, this study establishes a comprehensive quality assessment system tailored to architectural component scenarios. The framework includes efficiency metrics, model-related metrics, and subjective evaluations of visual quality and detail preservation. This multi-dimensional evaluation scheme validates the applicability of the proposed workflow across heritage conservation, documentation, and rapid modelling scenarios. Efficiency metrics. To achieve low cost and broad applicability on standard laptop hardware, the average processing time for dense reconstruction per image was approximately 40 seconds. For typical architectural components, 50 to 300 images were acquired, resulting in a total workflow duration of about 30–180 minutes. These metrics were used to evaluate the practical feasibility of the workflow under low-resource conditions. Model-related metrics. These metrics comprehensively evaluate model performance in terms of usability, data transfer ease, visualization quality, and post-processing flexibility. Such evaluation ensures the reconstructed models achieve functional efficiency while providing a smooth user experience and flexible data management in practical applications. The exported model files typically range from 30 to 500 MB, making them suitable for online platform display, transfer via messaging or email, mobile visualization, and integration into heritage conservation databases. Visual quality assessment. To address the limitations of traditional geometric metrics in capturing visual quality, this study introduces a subjective evaluation scheme focused on the preservation of texture and geometric details. While conventional geometric metrics (e.g., reprojection error) are widely used, they often fail to quantify perceptual degradations such as texture aliasing or topological noise. Accordingly, a standardized expert-based visual assessment is employed as perceptual ground truth to calibrate the effectiveness of the MTL indicator. Specifically, it examines trends in visual perception across different Mean Track Length ranges, validating how variations in reconstruction stability are reflected at the perceptual level. The visual assessment comprises the following dimensions: Edge continuity. Whether carved edges are continuous and intact, and whether the directional flow of wood grain appears smooth and coherent. Detail representation. Whether variations in carving depth are rendered naturally and whether fine-scale surface relief is adequately preserved. Surface roughness and noise aggregation. Whether high-frequency noise introduced during point cloud fusion, often resulting from accumulated redundant observations, leads to localized surface roughness or clustering of noise artifacts. The visual assessment used a five-point anchored rating scale (1 to 5), with each level corresponding to explicit quality criteria. A score of 1 indicates severe defects, making the model unsuitable for analysis or presentation. A score of 2 means only local features are discernible and overall quality is insufficient for practical use. A score of 3 indicates basic recognizability but evident fragmentation. A score of 4 represents a generally usable model with continuous major textures and geometric details, and minor defects that do not affect application. A score of 5 denotes excellent model quality with clear textures and well-preserved details. The final visual score was calculated as the average of the three visual dimensions. To ensure consistency, scores were averaged across ratings from five professionals in architectural heritage conservation (each with over 3 years of experience), with outlier ratings—defined as those deviating significantly from the group mean—excluded to mitigate subjective bias. The scoring examples are shown in Fig. 2 . 3. Experiment and Application: architectural heritage component 3D reconstruction To validate the feasibility of the proposed lightweight, low-cost 3D modelling workflow, this study conducted experiments on eight types of architectural components differing in size, texture strength, geometric configuration, and colour richness. All experiments were carried out using standard consumer-grade hardware, without professional photography equipment or high-performance computing platforms, and all software tools used in the workflow were freely available. All experiments were performed on a MECHREVO Aurora X (standard edition) laptop, featuring a 13th-generation Intel Core i7 processor, an NVIDIA GeForce RTX 5060 Laptop GPU with 8 GB of dedicated memory, and 16 GB of system RAM. Images were captured using a Redmi K80 smartphone with a 50 MP primary camera. All photos were taken in 1× mode, with an equivalent focal length of 24 mm and a resolution of 4096 × 3072 pixels. To ensure consistency, the focal length was fixed, exposure and white balance were locked, and HDR was disabled throughout image acquisition. Point cloud reconstruction was conducted with COLMAP version 3.12.6 (commit 4d5b60e), and mesh reconstruction as well as texture mapping were completed using MeshLab version 2025.07. 3.1. Components visualization and comparison To further validate the applicability of the proposed modelling workflow across diverse component types and conditions, experiments were conducted on eight representative categories of architectural components with varying scales, texture characteristics, and geometric complexities. The selected components encompass typical scenarios such as low-texture, small-scale elements; high-texture, large-scale elements; and components with intricate geometric features, as summarized in Table 2 . Table 2 Overview of basic characteristics, validation objectives, and model information of the reconstructed architectural components. ID Type Geometry Validation Objective Time(min) Size(MB) Result C1 Tibetan tower exterior wall Low Large-scale, low-texture applicability 88 196 Usable C2 Tibetan-style canvas Low Small-scale, high-colour applicability 9.3 140 Usable C3 Screen wall base Medium Large-scale, high-colour applicability 218 617 Usable C4 Wooden doors and windows Medium Small-scale, wooden component applicability 65 109 Usable C5 Wooden component High Large-scale, wooden component applicability 182 2070 Usable C6 Stone carving component High Small-scale, stone carving applicability 43 43 Usable C7 Stone carving component High Large-scale, stone carving applicability 43 53 Usable C8 Architectural balcony Low Simple geometric structure applicability 34 128 Usable The modelling results demonstrate that the proposed workflow consistently reproduces the principal geometric forms and texture features of architectural components across various types and characteristic conditions. As shown in Table 2 , the resulting models meet the requirements for component-level digital documentation under different component scenarios. Figure 3 presents a comparison between the corresponding on-site photographs and the reconstructed three-dimensional models. 3.2. Threshold determination In this experiment, the upper and lower modelling thresholds for each architectural component were established by analyzing the relationship between the number of input images and the corresponding Mean Track Length values during reconstruction. The upper threshold indicates the point at which models display clear, complete texture representation and well-defined geometric structures, while the lower threshold represents the minimum conditions required to preserve basic texture appearance and geometric integrity. From a photogrammetric perspective, Mean Track Length characterizes the average co-visibility of three-dimensional points across multiple views, serving as a comprehensive indicator of observational redundancy, geometric constraint stability, and image network connectivity. A low Mean Track Length means that 3D points are supported by only a few viewing rays, rendering the bundle adjustment problem ill-posed and highly susceptible to outliers[19]. As Mean Track Length increases, multi-view redundancy is enhanced, significantly improving the stability of the geometric solution during bundle adjustment. However, beyond a certain threshold, the marginal utility of additional observations declines (i.e., diminishing returns). Excessive redundancy fails to yield geometric improvements and instead introduces high-frequency noise and matching ambiguities. Therefore, Mean Track Length theoretically has a reasonable effective range, providing a photogrammetric basis for threshold-based optimization of image acquisition strategies. A total of 66 modelling experiments were conducted on various types of architectural components. The corresponding experimental data are summarized in Table 3 , and representative modelling results under different threshold conditions are shown in Fig. 4 . Table 3 Evaluation of modelling performance and threshold determination across MTL ranges for Components 1–8 ID MTL range Detail representation Edge continuity Surface roughness and noise Overall visual score Quality assessment Upper threshold Lower threshold C1 < 3.84 Incomplete Discontinuous Absent 4.26 Complete Continuous Present 5.0 Saturated C2 < 3.46 Incomplete Discontinuous Absent 4.29 Complete Continuous Present 5.0 Saturated C3 < 3.59 Incomplete Discontinuous Absent 4.20 Complete Continuous Present 5.0 Saturated C4 < 3.71 Incomplete Discontinuous Absent 4.25 Complete Continuous Present 5.0 Saturated C5 < 3.75 Incomplete Discontinuous Absent 4.19 Complete Continuous Present 5.0 Saturated C6 < 3.54 Incomplete Discontinuous Absent 4.11 Complete Continuous Present 5.0 Saturated C7 < 3.63 Incomplete Discontinuous Absent 4.24 Complete Continuous Present 5.0 Saturated C8 < 3.67 Incomplete Discontinuous Absent 4.19 Complete Continuous Present 5.0 Saturated As shown in the figure, the upper threshold results displayed in the first and third rows preserve overall geometric integrity but increasingly exhibit redundant noise. In contrast, the lower threshold results in the second and fourth rows are characterized by geometric instability or loss of fine details. 3.3. Mean Track Length (MTL) calibration As shown in Fig. 5 , despite clear differences among architectural components in scale, texture richness, and geometric complexity, the corresponding lower and upper threshold values exhibit strong clustering in their numerical distributions, with only minor fluctuations. Specifically, the lower thresholds of the tested components are mainly distributed within the range of approximately 3.5–3.8, while the upper thresholds are concentrated around 4.1–4.3, without significant dispersion or outliers. Based on these distribution patterns, the experimental results adopt 3.60–4.20 as a representative Mean Track Length interval for component-level modelling, summarizing the clustered range observed across different component experiments. 4. Discussion In SfM-based image-based 3D reconstruction research, various metrics are available for evaluating reconstruction quality, such as Mean Track Length, point cloud density, reprojection error, and subjective visual assessment[20]. However, for component-level modelling of traditional architecture, most of these indicators serve as post-hoc evaluation results and cannot directly reflect the stability of multi-view geometric constraints or the impact of image redundancy on the reconstruction process. In this study, 66 practical modelling experiments were conducted on eight traditional architectural components with varying sizes, texture characteristics, and geometric complexities. The corresponding sparse reconstruction statistics for each experiment were systematically extracted. Using Fig. 6 as an example, the mean reprojection error remains consistently low as the number of images for each component increases. This is primarily because reprojection error serves as an optimization target during bundle adjustment. The solver minimizes residuals even when the underlying geometry is weak or ill-posed, provided that the feature correspondences are mathematically consistent. Consequently, low reprojection error does not necessarily guarantee structural stability. Further analysis reveals that although the Number of Observations rises with the number of images, its absolute values differ greatly among components due to variations in scale, texture, and coverage, making it unsuitable as a unified criterion for assessing modelling quality. In contrast, Mean Track Length offers a more consistent measure of the reliability of multi-view geometric constraints and exhibits stable variation patterns across different components. When Mean Track Length is low, reconstructed models frequently suffer from insufficient geometric constraints, leading to local structural instability, geometric collapse, and texture bleeding. Conversely, when Mean Track Length surpasses a certain range, even as the number of observations continues to rise, improvements in model quality tend to plateau, and excessive redundancy may introduce increased noise and surface roughness. Notably, cross-component comparisons show that Mean Track Length exhibits strong correlation and consistency across various components, further confirming its value as a process-oriented indicator for component-level image-based 3D modelling. Building upon the above analysis of indicator rationality, this study further defines an operational threshold range for Mean Track Length, which provides important engineering value for guiding image acquisition density control and reconstruction termination decisions. Within the COLMAP/SfM-based image-based 3D reconstruction framework, there has historically been a lack of stable intermediate indicators to guide image acquisition density and termination decisions during modelling. Most existing studies rely on point cloud size, reprojection error, or subjective visual assessment—metrics that are primarily post hoc and offer limited process-level guidance. Given the established validity of Mean Track Length as a core indicator for component-level modelling, further defining its upper and lower thresholds is of significant engineering importance for optimizing image acquisition strategies and the reconstruction workflow. Grounded in the stability of multi-view geometric constraints, this study introduces Mean Track Length as the core evaluation indicator for component-level modelling quality[21]. Drawing on 66 modelling experiments and analysis of 462 datasets, an operational Mean Track Length threshold range (3.60–4.20) is proposed for traditional architectural components. Results show that when Mean Track Length drops below 3.60, stable geometric constraints are difficult to achieve, and the resulting models frequently present issues such as texture bleeding and local geometric collapse. When Mean Track Length exceeds 4.20, further increases in image redundancy fail to improve reconstruction quality and instead introduce redundant effects, including noise accumulation, point cloud redundancy, and increased reprojection error. Therefore, a Mean Track Length range of 3.60–4.20 can be considered an optimal interval that balances modelling quality, computational efficiency, and cost control. This study demonstrates that, in component-level image-based modelling, blindly increasing image redundancy does not lead to a linear improvement in reconstruction reliability; instead, it may introduce unnecessary computational burdens and decision-making risks. This threshold transforms what was once an experience-dependent modelling workflow into a threshold-driven, quantifiable, and reproducible technical process. It offers clear criteria for image acquisition strategies and modelling decisions in the reconstruction of traditional architectural components, shifting the process from empirical methods to a controllable, threshold-based workflow. Consequently, it establishes a process control framework for component-level image-based modelling of architectural heritage that is grounded in explicit physical meaning and practical engineering applicability. From a methodological perspective, the threshold-driven modeling strategy proposed in this study differs in its technical paradigm from the encapsulated photogrammetric solutions represented by commercial software. Existing commercial solutions are generally designed with encapsulated workflows to prioritize automation and user-friendliness. While efficient, this approach limits user intervention. In contrast, our proposed method offers process-level transparency, allowing for the explicit monitoring of geometric stability that is critical for scientific documentation. In contrast, this study develops a component-level modelling workflow using the open-source SfM framework COLMAP, with a core advantage in the interpretability and controllability of the reconstruction process. By introducing intermediate indicators such as Mean Track Length, the workflow enables continuous monitoring of feature matching performance and 3D geometric stability during reconstruction. This allows for real-time assessment of whether current image acquisition is within the “effective information range” and supports data-driven decisions on continuing or terminating image acquisition. The threshold-based (3.60–4.20) process control mechanism transforms modelling failures from opaque outcomes into quantifiable, analysable, and correctable process-level issues. In summary, commercial modelling software and the approach proposed in this study represent two distinct technical paradigms, each tailored to different objectives rather than a straightforward comparison of superiority. Commercial solutions prioritize automation and result-oriented efficiency, while the proposed approach emphasizes process transparency, interpretability, and methodological reproducibility. For applications such as traditional architectural documentation, heritage surveys, and academic research, the threshold-driven modelling method introduced in this study offers a more controllable and verifiable technical alternative for component-level reconstruction. During the analysis of the experimental data, a rather unexpected phenomenon was also observed. Existing studies generally suggest that texture-sparse or low-texture image regions increase the difficulty of local feature-based matching, thereby adversely affecting the quality of SfM reconstruction, as the number of extractable features and the reliability of feature matching are typically lower than those of high-texture components[22]. However, as shown in Fig. 7 , no significant difference is observed in the lower thresholds between low-texture (C1) and high-texture (C6) components. This phenomenon may be attributed to the combined effects of multiple filtering processes and geometric compensation mechanisms during sparse reconstruction. On the one hand, in low-texture regions, although the number of extractable feature points is limited, those that remain after geometric consistency verification and bundle adjustment are typically more stable structural features, which tend to exhibit relatively longer track lengths, thereby statistically elevating the Mean Track Length of low-texture components[23]. On the other hand, some low-texture components simultaneously possess well-defined geometric structures, such as edges, corners, or component contours. These geometric constraints can partially compensate for the lack of texture information, making the effectiveness of multi-view geometric constraints comparable to that of high-texture components[24]. To address the difficulty of point cloud extraction under low-texture conditions, this study further explored improvement strategies such as masking and secondary re-injection. However, these methods were only investigated as exploratory attempts and were not incorporated into the threshold-driven modelling framework proposed in this work. Although this study proposes a Mean Track Length–based threshold range for component-level image-based 3D modelling based on multiple component experiments, several limitations remain. First, the experimental samples are primarily drawn from eight representative architectural component types. While these samples offer diversity in texture richness, geometric complexity, and scale, the overall sample size is still relatively limited. Second, the proposed threshold framework is mainly designed for component-level image-based 3D modelling; its applicability to large-scale building or block-level modelling tasks requires further validation in future research. Additionally, as a statistical indicator derived from the sparse reconstruction stage, Mean Track Length (MTL) can be influenced by factors such as image resolution, feature extraction parameters, and image acquisition trajectories. Although this study maintained consistent parameter settings and modelling procedures throughout the experiments, the proposed threshold range may still vary under different hardware conditions or reconstruction configurations. Therefore, in practical applications, limited preliminary testing and calibration are recommended to adapt the threshold to specific workflows. Finally, the determination of the thresholds was based on a combined consideration of statistical indicators obtained from the sparse reconstruction stage and the actual modelling performance. However, the evaluation of model quality still lacks no-reference image quality assessment (No-Reference IQA) metrics from the perspective of image signal processing (ISP) to enable the automation of this process. The introduction and development of such metrics will therefore constitute an important direction for future work. Nevertheless, the judgment process in this study is constrained by the overall trends observed across multiple experimental results, thereby ensuring the rationality and interpretability of the derived thresholds within the scope of this research. 5. Conclusions and Prospects This study proposes a transferable, stability-threshold-centered decision-making approach for component-level image-based modelling. Designed to be low-cost, reproducible, and stable on standard computing devices, the workflow is intended for scenarios such as digital documentation of traditional architectural components and grassroots heritage surveys. Its engineering feasibility and process stability are demonstrated through experimental validation. Building on this workflow, the study introduces Mean Track Length as a threshold indicator for component-level image-based modelling and, through multiple experimental validations, establishes an operational effective range (3.60–4.20). By establishing explicit, quantifiable criteria for acquisition density, this framework transforms component-level photogrammetry from an empirical, experience-dependent practice into a controllable, engineering-grade workflow. Future research can be pursued in several directions. Algorithmically, deeper exploration of the relationship between Mean Track Length and the internal multi-view geometric constraint mechanisms of SfM could provide a theoretical foundation for the threshold values proposed. In terms of application, the threshold-driven modelling strategy developed for component-level reconstruction in this study could be extended to broader heritage digitization scenarios, such as building façades, individual structures, and even urban block–scale contexts. At the indicator system level, Mean Track Length may be integrated with additional metrics, such as texture strength and viewpoint distribution, to establish a more comprehensive evaluation and decision-making system for component-level image-based 3D modelling quality. This would offer robust technical support for the digitalization of architectural heritage at larger scales and under more complex conditions. Declarations Acknowledgement This work is supported by China-Portugal Joint Laboratory of Cultural Heritage Conservation Science (No. SDYY2405), Sichuan Science and Technology Program (Grant Number: 2025ZNSFSC1309), and the Fundamental Research Funds for the Central Universities, Southwest Minzu University (No. 2024SYJSCX147, No. ZYN2025069). Declaration of Conflicting Interests The authors declare no potential conflicts of interest concerning the research, authorship, and/or publication of this article. Availability of data and materials The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Author Contributions Statement LW conceived the research idea, designed the overall methodology, conducted the photogrammetric experiments, performed data analysis, and drafted the original manuscript; YD contributed to data acquisition and image preprocessing, and assisted with experimental implementation and result validation; JJ participated in the development of the modelling workflow and supported data organization and visualization; YL assisted with experimental data collection and contributed to the evaluation of reconstruction results; HZ provided support in software operation, parameter testing, and preliminary result inspection; QY supervised and coordinated the research and reviewed and revised the manuscript; YZ provided academic guidance on research design, contributed to methodological refinement, and reviewed the manuscript. All authors reviewed the manuscript. References Liu S, Bin Mamat MJ. Application of 3D laser scanning technology for mapping and accuracy assessment of the point cloud model for the Great Achievement Palace heritage building. Herit Sci. 2024;12(1):153. Ma H, Zhou Z, Wang Y. Digital preservation and development of architectural heritage from a virtual perspective: a systematic review. npj Herit Sci. 2025;13(1):594. Buragohain D, Meng Y, Deng C, Li Q, Chaudhary S. Digitalizing cultural heritage through metaverse applications: challenges, opportunities, and strategies. Herit Sci. 2024;12(1):295. Zheng H, Chen L, Hu H, Wang Y, Wei Y. Research on the Digital Preservation of Architectural Heritage Based on Virtual Reality Technology. Buildings. 2024;14(5):1436. García-Molina DF, González-Merino R, Rodero-Pérez J, Carrasco-Hurtado B. 3D documentation for the conservation of historical heritage: the Castle of Priego de Córdoba (Spain). Virtual Archaeol Rev. 2021;12(24):115–30. He J, Jia Q. Opti3D for low light enhancement and calibration free 3D digitization of cultural relics. npj Herit Sci. 2025;13(1):357. Stylianidis E, Evangelidis K, Vital R, Dafiotis P, Sylaiou S. 3D Documentation and visualization of cultural heritage buildings through the application of geospatial technologies. Heritage. 2022;5(4):2818–32. Liu J, Li B. Terrestrial laser scanning (TLS) survey and building information modeling (BIM) of the Edmund Pettus Bridge: A case study. Int Archives Photogrammetry Remote Sens Spat Inform Sci. 2024;48:379–86. Rahaman H, Champion E. To 3D or not 3D: Choosing a photogrammetry workflow for cultural heritage groups. Heritage. 2019;2(3):1835–51. Marčiš M, Fraštia M, Terao Vošková K. Potential of low-cost UAV photogrammetry for documenting hard-to-access interior spaces through building openings. Heritage. 2024;7(11):6173–91. Neumann KA, Tausch R, Kutlu H, Kuijper A, Santos P, Fellner D. Point cloud quality metrics for incremental image-based 3D reconstruction. Multimedia Tools Appl. 2025;84(32):39123–41. Calisi D, Botta S, Cannata A. Integrated surveying, from laser scanning to UAV systems, for detailed Documentation of architectural and archeological heritage. Drones. 2023;7(9):568. Klapa P, Żygadło A, Pepe M. 3D heritage reconstruction through HBIM and multi-source data fusion: geometric change analysis across decades. Appl Sci. 2025;15(16):8929. Maiwald F, Feurer D, Eltner A. Solving photogrammetric cold cases using AI-based image matching: New potential for monitoring the past with historical aerial images. ISPRS J Photogrammetry Remote Sens. 2023;206:184–200. Schonberger JL, Frahm J-M. Structure-from-motion revisited. In: Proceedings of the IEEE conference on computer vision and pattern recognition : 2016; 2016: 4104–4113. Cignoni P, Callieri M, Corsini M, Dellepiane M, Ganovelli F, Ranzuglia G. Meshlab: an open-source mesh processing tool. In: Eurographics Italian chapter conference: 2008 : Salerno; 2008: 129–136. Pataki Z, Sarlin P-E, Schönberger JL, Pollefeys M. MP-SfM: Monocular Surface Priors for Robust Structure-from-Motion. In: Proceedings of the Computer Vision and Pattern Recognition Conference : 2025; 2025: 21891–21901. Heim RH, Okole N, Steppe K, Van Labeke M-C, Geedicke I, Maes WH. An applied framework to unlocking multi-angular UAV reflectance data: a case study for classification of plant parameters in maize (Zea mays). Precision Agric. 2024;25(3):1751–75. Cueto Zumaya CR, Catalano I, Queralta JP. Building Better Models: Benchmarking Feature Extraction and Matching for Structure from Motion at Construction Sites. Remote Sens. 2024;16(16):2974. Hou Q, Xia R, Zhang J, Feng Y, Zhan Z, Wang X. Learning visual overlapping image pairs for SfM via CNN fine-tuning with photogrammetric geometry information. Int J Appl Earth Obs Geoinf. 2023;116:103162. Huang D, Qin R, Elhashash M. Bundle adjustment with motion constraints for uncalibrated multi-camera systems at the ground level. ISPRS J Photogrammetry Remote Sens. 2024;211:452–64. Yang X, Jiang G. A practical 3D reconstruction method for weak texture scenes. Remote Sens. 2021;13(16):3103. Lee J, Yoo S. Dense-SfM: Structure from Motion with Dense Consistent Matching. In: Proceedings of the Computer Vision and Pattern Recognition Conference : 2025; 2025: 6404–6414. Zhao L, Guo F, Zhu Y, Wang H, Zhou B. A Generalized Voronoi Diagram-Based Segment-Point Cyclic Line Segment Matching Method for Stereo Satellite Images. Remote Sens. 2024;16(23):4395. Additional Declarations No competing interests reported. Supplementary Files SupplementaryScriptforFrameSampling.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 07 Mar, 2026 Reviews received at journal 05 Mar, 2026 Reviews received at journal 24 Feb, 2026 Reviewers agreed at journal 09 Feb, 2026 Reviewers agreed at journal 08 Feb, 2026 Reviewers invited by journal 08 Feb, 2026 Editor assigned by journal 03 Feb, 2026 Submission checks completed at journal 01 Feb, 2026 First submitted to journal 27 Jan, 2026 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-8709145","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":589060081,"identity":"ce700b1a-0c61-432d-bf71-f43975274c05","order_by":0,"name":"Luo Wang","email":"","orcid":"","institution":"Southwest Minzu University","correspondingAuthor":false,"prefix":"","firstName":"Luo","middleName":"","lastName":"Wang","suffix":""},{"id":589060082,"identity":"9c36eef1-99ff-4f17-b8c3-0b25893d16d8","order_by":1,"name":"Yating Duan","email":"","orcid":"","institution":"Southwest Minzu University","correspondingAuthor":false,"prefix":"","firstName":"Yating","middleName":"","lastName":"Duan","suffix":""},{"id":589060083,"identity":"7ba163b6-9e8d-4731-8ad1-eb9cbb597778","order_by":2,"name":"Jungang Jiang","email":"","orcid":"","institution":"Southwest Minzu University","correspondingAuthor":false,"prefix":"","firstName":"Jungang","middleName":"","lastName":"Jiang","suffix":""},{"id":589060084,"identity":"a137e2b3-5ce7-4878-9f6d-90002d89f8c6","order_by":3,"name":"Yiheng Liu","email":"","orcid":"","institution":"Southwest Minzu University","correspondingAuthor":false,"prefix":"","firstName":"Yiheng","middleName":"","lastName":"Liu","suffix":""},{"id":589060085,"identity":"692febcb-39eb-4491-8531-1f4bf9acbf67","order_by":4,"name":"Hailin Zhang","email":"","orcid":"","institution":"Southwest Minzu University","correspondingAuthor":false,"prefix":"","firstName":"Hailin","middleName":"","lastName":"Zhang","suffix":""},{"id":589060086,"identity":"d57599ae-a995-4238-a066-a2ecd15b0835","order_by":5,"name":"Qianru Yang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA70lEQVRIiWNgGAWjYNACAxDBfADKSyBaCxtMKVFawIDHgDgtBsfPHn7xpsAuz3xGzufPvG13GPjZcwwYfu7Ao+VMXprlHIPkYpkbudukedueMUj2vDFg7D2DR8uBHDNjHgPmxBkSuduYc9sOMxjcyDFgZmzDo+X8G5CWeqCWnMefQVrsCWq5kWP8mMfgMEgLgzTYFgkCWiRvvDFjnGNwPHEGzzMz6T/nDvNInHlWcLAXjxa+8znGH978qU6cwZ78+OOMssNy/O3JGx/8xKNF4QADmwQPkgCYfQC3BgYG+QYG5g88+FSMglEwCkbBKAAAJj9Sdj8ur5YAAAAASUVORK5CYII=","orcid":"","institution":"Southwest Minzu University","correspondingAuthor":true,"prefix":"","firstName":"Qianru","middleName":"","lastName":"Yang","suffix":""},{"id":589060087,"identity":"d6199e23-e85f-4695-80fd-4d402e1d7551","order_by":6,"name":"Yin Zhang","email":"","orcid":"","institution":"Southwest Minzu University","correspondingAuthor":false,"prefix":"","firstName":"Yin","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2026-01-27 10:30:14","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8709145/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8709145/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102962225,"identity":"dba4c0a7-dda5-481e-af34-a68dcae2dce3","added_by":"auto","created_at":"2026-02-19 04:05:42","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":312446,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the proposed low-cost architectural heritage modelling framework based on quantitative threshold control.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/9d7c40eeb1091dd90e7935b9.png"},{"id":102503676,"identity":"1aeb87ba-0062-4a9e-a0cf-59c62a16e348","added_by":"auto","created_at":"2026-02-12 11:05:39","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":79471,"visible":true,"origin":"","legend":"\u003cp\u003eRepresentative examples of component-level 3D reconstruction results corresponding to different visual quality scores (1–5).\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/933f3e838e6c48e4e694be04.png"},{"id":102746821,"identity":"024b479a-aa8e-4f82-b22a-ebf1c8a9cfa8","added_by":"auto","created_at":"2026-02-16 09:01:41","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1199990,"visible":true,"origin":"","legend":"\u003cp\u003eComparison between on-site photographs and the corresponding reconstructed models of architectural components, including: Component 1 on-site photograph (a) and reconstructed model (e), Component 2 on-site photograph (b) and reconstructed model (f), Component 6 on-site photograph (c) and reconstructed model (g), Component 7 on-site photograph (d) and reconstructed model (h), Component 5 on-site photograph (i) and reconstructed model (m), Component 3 on-site photograph (j) and reconstructed model (n), Component 4 on-site photograph (k) and reconstructed model (o), and Component 8 on-site photograph (l) and reconstructed model (p).\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/20bef0df546481719ea1cee7.png"},{"id":102503679,"identity":"6957c08b-1fc8-4548-b7c5-acc65a8c320c","added_by":"auto","created_at":"2026-02-12 11:05:40","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":385393,"visible":true,"origin":"","legend":"\u003cp\u003eComparison of upper and lower threshold effects for each component, including: Component 1 upper threshold effect (a) and lower threshold effect (e), Component 2 upper threshold effect (b) and lower threshold effect (f), Component 6 upper threshold effect (c) and lower threshold effect (g), Component 7 upper threshold effect (d) and lower threshold effect (h), Component 5 upper threshold effect (i) and lower threshold effect (m), Component 3 upper threshold effect (j) and lower threshold effect (n), Component 4 upper threshold effect (k) and lower threshold effect (o), and Component 8 upper threshold effect (l) and lower threshold effect (p).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/da53f2bf70094e678ed86047.png"},{"id":102747016,"identity":"57641f52-124a-4a74-b0a5-332d66c41581","added_by":"auto","created_at":"2026-02-16 09:03:34","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":31606,"visible":true,"origin":"","legend":"\u003cp\u003eUpper and lower threshold distributions of Mean Track Length (MTL) for different architectural components. Left: upper threshold values; right: lower threshold values.\u003c/p\u003e","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/4f870369ae52d1e722dc5d61.jpeg"},{"id":102503680,"identity":"0996ac6a-ed56-4a10-94a3-00d9eb24ae88","added_by":"auto","created_at":"2026-02-12 11:05:40","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":696615,"visible":true,"origin":"","legend":"\u003cp\u003eRelationships between the number of images and Mean Track Length, Number of Observations, and mean reprojection error for different architectural components.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/e71300e8fd8f7c2d805effca.png"},{"id":102503682,"identity":"21777566-dc8c-45c2-83c7-c34f27496886","added_by":"auto","created_at":"2026-02-12 11:05:40","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":185568,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of lower Mean Track Length thresholds for eight architectural components (C1–C8).\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/5a92970ad593fbf4dac0f057.png"},{"id":103056438,"identity":"3c0c345e-4241-40d5-a807-9a2cc5af95ab","added_by":"auto","created_at":"2026-02-20 09:10:26","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4170165,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/f3eec572-ea0d-417c-bd76-7fce5541473f.pdf"},{"id":102503678,"identity":"83528faa-018f-4604-a0a8-6bf1d4657ed4","added_by":"auto","created_at":"2026-02-12 11:05:40","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":23223,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryScriptforFrameSampling.docx","url":"https://assets-eu.researchsquare.com/files/rs-8709145/v1/c766aa815f1cf4d438c5e343.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Novel Mean Track Length-Driven 3D Framework via COLMAP for Architectural Heritage Photogrammetry","fulltext":[{"header":"1. Introduction","content":"\u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e1.1. Research background\u003c/h2\u003e \u003cp\u003eAs pressures on cultural heritage conservation continue to increase, digital technologies\u0026mdash;particularly three-dimensional (3D) modelling, point cloud generation, and virtual visualization\u0026mdash;have played an increasingly important role in heritage documentation, conservation, and management[1]. Previous studies have shown that 3D digital technologies can generate highly detailed digital heritage models and provide essential baseline data to support long-term preservation, restoration planning, and public education[2]. Moreover, digital heritage practices are progressively extending toward virtual environments and public engagement, enhancing the global accessibility and interpretability of cultural heritage through digital platforms[3]. Empirical studies conducted in China likewise highlight the significance of digital technologies in the conservation of heritage buildings, emphasizing their advantages in condition recording, decision support, and the development of interactive heritage experiences[4].\u003c/p\u003e \u003cp\u003eWith the continuous advancement of cultural heritage conservation and digital technologies, the digital documentation of architectural heritage has become an important research focus in the fields of architecture, archaeology, and heritage conservation[5]. Digital technologies\u0026mdash;particularly image-based three-dimensional (3D) modelling\u0026mdash;enable non-contact and accurate recording of the geometric forms, surface details, and texture characteristics of heritage assets, providing critical support for long-term preservation, restoration planning, and virtual presentation[6]. Through digital methods, architectural heritage can be virtually reconstructed and systematically archived, while precise 3D data can be generated to support restoration interventions, effectively reducing the risk of damage associated with physical contact[7].\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e1.2. Literature review\u003c/h2\u003e \u003cp\u003eExisting approaches to three-dimensional documentation of architectural heritage can be broadly categorized into three main pathways. Three-dimensional laser scanning enables rapid acquisition of high-density, high-precision point cloud data and offers clear advantages in capturing the overall geometry of large-scale buildings and complex spatial environments. However, the high cost of equipment, along with demanding operational and post-processing requirements create significant barriers to entry, thereby stifling its widespread adoption in routine heritage surveys[8]. Image-based photogrammetric techniques reconstruct three-dimensional models from multi-view imagery without relying on dedicated surveying equipment, offering a comparatively low-cost solution. As a result, photogrammetry has become one of the most widely adopted approaches in digital heritage research. Nevertheless, many mature software platforms and algorithmic frameworks are characterized by steep learning curves or high commercial licensing costs, which continue to limit their accessibility and broader application[9]. Unmanned aerial vehicle (UAV)-based image acquisition and oblique photogrammetry are well suited for the overall documentation of large-scale heritage assets. However, these approaches are primarily restricted to outdoor environments and impose additional requirements related to flight control and airspace regulation, which continue to limit their applicability in certain heritage documentation contexts[10].\u003c/p\u003e \u003cp\u003eAlthough structure-from-motion (SfM) techniques offer clear advantages in terms of cost efficiency and equipment requirements, existing studies have predominantly focused on post-reconstruction model accuracy assessment or algorithmic improvements. This reliance on post-hoc evaluation offers limited decision support during the acquisition phase, as it lacks real-time, process-oriented quantitative metrics essential for preempting reconstruction failure. Crucially, conventional metrics like reprojection error often remain low even in geometrically unstable models due to solver optimization, masking underlying data sufficiency issues. Often, by the time quality issues are detected in post-processing, the field team has already left the heritage site, making re-acquisition impossible. In practical applications, this deficiency often leads to reconstruction failures caused by insufficient image acquisition, or, conversely, to excessive data redundancy and increased computational costs due to over-acquisition. Therefore, there is a pressing need to introduce quantitative indicators that provide quality feedback during the reconstruction process, enabling dynamic regulation of image acquisition and early prediction of modelling reliability[11].\u003c/p\u003e \u003cp\u003eIn addition, the integration of multiple techniques\u0026mdash;such as laser scanning, photogrammetry, and UAV-based acquisition\u0026mdash;has become an emerging trend in comprehensive digital heritage documentation. However, such integrated approaches are typically associated with higher overall costs and increased workflow complexity[12]. Although numerous studies have demonstrated the respective advantages of these technologies in heritage reconstruction, most existing research has focused primarily on macro-scale architectural structures, overall scenes, or large heritage sites. At the component and detail levels, there is still a lack of targeted threshold definitions, standardized workflows, and low-cost, scalable solutions for high-precision modelling[13]. Moreover, substantial barriers related to resources, equipment, and professional expertise remain, further constraining the widespread adoption and practical dissemination of these techniques.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e1.3. Research recaps and gaps\u003c/h2\u003e \u003cp\u003eCompared with other open-source photogrammetric tools, COLMAP demonstrates notable advantages in the robustness of feature extraction and matching, the modelling of multi-view geometric constraints, and the accessibility of process-oriented metrics. These characteristics enable COLMAP not only to support 3D model generation, but also to facilitate process-oriented analysis of reconstruction stability, making it particularly suitable for component-level image-based 3D reconstruction research[14].\u003c/p\u003e \u003cp\u003eBuilding upon this, this study establishes a low-cost, open-source modelling workflow centered on COLMAP and MeshLab[15, 16]. Instead of pursuing the absolute highest geometric accuracy, this workflow prioritizes the cost-effectiveness ratio and process reliability for rapid heritage documentation. Based on extensive comparative experiments, and considering variations in component size, texture strength, geometric structure, and color characteristics, a set of practical modelling threshold ranges is proposed to assist in balancing image redundancy, reconstruction stability, and modelling efficiency, thereby avoiding redundancy effects and reducing unnecessary data acquisition and computational costs[17].\u003c/p\u003e \u003cp\u003eIn addition, a frame extraction strategy is introduced to simplify the data acquisition process, making image capture and modelling workflows more controllable and providing a threshold-driven, risk-aware decision pathway for component-level image-based modelling. This study aims to promote the application of lightweight, low-cost, and reproducible digital documentation methods for architectural heritage in grassroots surveys and educational contexts, offering quantitative references for determining when image-based modelling becomes sufficiently reliable at the component level, and supporting the standardization and methodological development of heritage digitization technologies.\u003c/p\u003e \u003c/div\u003e"},{"header":"2. Methodology","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Research workflow\u003c/h2\u003e \u003cp\u003eThis study introduces a low-cost, easily reproducible, and parameter-fixed open-source workflow designed to provide a practical and affordable digital modelling solution for traditional architectural components. To clearly illustrate the proposed workflow, the overall digital modelling process is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. This figure presents the complete pipeline, including image acquisition and preprocessing, feature extraction, matching, and three-dimensional reconstruction, followed by mesh generation, texture mapping, and quality evaluation.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Data acquisition and preprocessing\u003c/h2\u003e \u003cp\u003eImage acquisition was performed using standard consumer-grade smartphones to capture photographs of architectural components. To ensure data completeness and operational efficiency, a multi-view acquisition strategy was employed[18], whereby components were densely imaged from multiple viewpoints along surrounding trajectories. This method guaranteed comprehensive coverage of each component while capturing sufficient texture information for reliable three-dimensional reconstruction. For components with complex textures and fine-scale details, particular attention was given to minimizing the impact of illumination variations on the reconstruction results. All components were photographed using the same smartphone to maintain consistency in image properties, including resolution, focal length, and aperture. Additionally, care was taken to preserve the original image metadata during data transfer and format conversion, thereby preventing the loss of critical information such as pixel resolution, focal length, and aperture settings.\u003c/p\u003e \u003cp\u003eTo ensure reproducibility in subset generation, a standardized sequential renaming protocol was applied to all valid images prior to ingestion into the reconstruction pipeline.\u003c/p\u003e \u003cp\u003eAfter the renaming procedure, a uniform frame extraction strategy was applied to the high-density image sets. Initially, images were subsampled by selecting files with sequential indices of 1, 3, 5, and so on, to form the first dataset. Additional datasets were then generated by extracting images with indices of 1, 4, 7, 11, etc., thereby producing multiple image subsets with varying quantities of images. The implementation code for the frame-sampling method is provided in the Supplementary Materials.\u003c/p\u003e \u003cp\u003eBy varying the sampling step size S, multiple image subsets containing different numbers of images were generated, allowing for analysis of the impact of image quantity on three-dimensional reconstruction quality. This frame extraction strategy facilitates the creation of multiple image datasets of varying sizes under consistent acquisition conditions, eliminating the need for repeated image collection. The validation results demonstrating the effectiveness of the frame extraction strategy are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eThis frame extraction strategy facilitates the creation of multiple image datasets of varying sizes without the need for repeated physical collection. Crucially, this approach allows for a controlled simulation of varying acquisition densities under identical illumination and sensor conditions, thereby isolating image quantity as the sole variable affecting reconstruction quality.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eValidation of the effectiveness of the frame extraction strategy. Note: \u0026ldquo;Saturated\u0026rdquo; is defined as the state where additional images increase file size and processing time without yielding any perceptible improvement in the Visual Quality Score (maintaining a score of 5.0).\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent ID\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFrame extraction strategy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of images\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMean Track Length\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eReconstruction quality\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eWithin effective range\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFull acquisition (no frame extraction)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e288\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e5.40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo (redundant)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eS\u0026thinsp;=\u0026thinsp;2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e144\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e4.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo (redundant)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eS\u0026thinsp;=\u0026thinsp;3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e4.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAcceptable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eS\u0026thinsp;=\u0026thinsp;4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e4.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAcceptable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eS\u0026thinsp;=\u0026thinsp;5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAcceptable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eS\u0026thinsp;=\u0026thinsp;6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAcceptable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eYes\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComponent 3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eS\u0026thinsp;=\u0026thinsp;7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eUnacceptable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Key parameter settings\u003c/h2\u003e \u003cp\u003eDuring feature extraction, matching, and three-dimensional reconstruction with COLMAP, default parameter settings were consistently used to enhance the reproducibility of the proposed workflow. However, in practical applications\u0026mdash;especially for low-texture components or image sets with limited overlap, where matching and reconstruction are more challenging\u0026mdash;certain parameters can be adjusted to improve feature acquisition and reconstruction robustness. The following parameter adjustments serve as empirical heuristics for challenging scenarios and act as a supplement to the core threshold-driven workflow.\u003c/p\u003e \u003cp\u003eDuring the feature extraction stage, the parameter max_image_size determines the target resolution to which images are uniformly rescaled before feature extraction. At lower resolutions, fewer feature points are typically detected, while increasing this value enables the extraction of more features but also raises computational demands. Similarly, max_num_features specifies the maximum number of feature points that can be extracted from each image; increasing this parameter can improve matching robustness by providing more features, but it also results in greater memory usage and longer processing times. Therefore, both parameters should be adjusted based on the characteristics of the image set and the available computational resources.\u003c/p\u003e \u003cp\u003eDuring the feature matching stage, the parameter max_num_matches sets the maximum number of feature correspondences established between each image pair. Increasing this value allows more matches to be retained, which may enhance matching completeness but also leads to greater memory consumption. Therefore, max_num_matches should be adjusted according to the specific characteristics of the image set and the available computational resources.\u003c/p\u003e \u003cp\u003eDuring mesh generation in MeshLab, considering both modelling quality and hardware constraints, only the Reconstruction Depth parameter was adjusted to a value of 15, while all other parameters remained at their default settings.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Data storage\u003c/h2\u003e \u003cp\u003eAfter completing dense point cloud reconstruction and subsequent cleaning, the point cloud data exported from COLMAP were imported into MeshLab for mesh reconstruction and texture mapping. First, a manual filtering process was conducted to remove background points unrelated to the architectural component, retaining only those corresponding to the target component. This step helps prevent redundant data from interfering with subsequent mesh generation and texture mapping procedures.\u003c/p\u003e \u003cp\u003eAfter point cloud cleaning, mesh reconstruction was performed. In MeshLab, a continuous triangular mesh surface was generated from the oriented point cloud using the menu path Filters \u0026rarr; Remeshing, Simplification and Reconstruction \u0026rarr; Surface Reconstruction: Screened Poisson. This method effectively suppresses the influence of point cloud noise on the mesh structure while preserving overall geometric continuity, thus providing a stable geometric foundation for subsequent texture mapping.\u003c/p\u003e \u003cp\u003ePrior to texture parameterization and mapping, representative original images intended for texture mapping were imported into MeshLab. To ensure high-quality texture mapping, the image dataset was selected to cover key component regions and capture typical texture characteristics. The selected images were loaded into the system via File \u0026rarr; Import Raster, providing the data source for subsequent image-based texture mapping procedures.\u003c/p\u003e \u003cp\u003eAfter mesh generation and image data import, texture parameterization and mapping were performed. First, under Filters \u0026rarr; Texture, the \u0026ldquo;Parameterization\u0026thinsp;+\u0026thinsp;texturing from registered rasters\u0026rdquo; function was used to complete texture mapping with the registered multi-view original images. The image information was projected onto the mesh surface, resulting in a textured model with a realistic visual appearance.\u003c/p\u003e \u003cp\u003eAfter completing mesh reconstruction and texture mapping, the model can be exported based on subsequent application requirements. MeshLab supports various common 3D model formats, such as .3ds, .obj, and .glb, offering strong compatibility and flexibility for further processing, visualization, and cross-platform applications.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.5. Evaluation and result validation\u003c/h2\u003e \u003cp\u003eTo systematically evaluate the effectiveness, stability, and scalability of the proposed lightweight modelling framework, this study establishes a comprehensive quality assessment system tailored to architectural component scenarios. The framework includes efficiency metrics, model-related metrics, and subjective evaluations of visual quality and detail preservation. This multi-dimensional evaluation scheme validates the applicability of the proposed workflow across heritage conservation, documentation, and rapid modelling scenarios.\u003c/p\u003e \u003cp\u003e \u003cb\u003eEfficiency metrics.\u003c/b\u003e To achieve low cost and broad applicability on standard laptop hardware, the average processing time for dense reconstruction per image was approximately 40 seconds. For typical architectural components, 50 to 300 images were acquired, resulting in a total workflow duration of about 30\u0026ndash;180 minutes. These metrics were used to evaluate the practical feasibility of the workflow under low-resource conditions.\u003c/p\u003e \u003cp\u003e \u003cb\u003eModel-related metrics.\u003c/b\u003e These metrics comprehensively evaluate model performance in terms of usability, data transfer ease, visualization quality, and post-processing flexibility. Such evaluation ensures the reconstructed models achieve functional efficiency while providing a smooth user experience and flexible data management in practical applications. The exported model files typically range from 30 to 500 MB, making them suitable for online platform display, transfer via messaging or email, mobile visualization, and integration into heritage conservation databases.\u003c/p\u003e \u003cp\u003e \u003cb\u003eVisual quality assessment.\u003c/b\u003e To address the limitations of traditional geometric metrics in capturing visual quality, this study introduces a subjective evaluation scheme focused on the preservation of texture and geometric details. While conventional geometric metrics (e.g., reprojection error) are widely used, they often fail to quantify perceptual degradations such as texture aliasing or topological noise. Accordingly, a standardized expert-based visual assessment is employed as perceptual ground truth to calibrate the effectiveness of the MTL indicator. Specifically, it examines trends in visual perception across different Mean Track Length ranges, validating how variations in reconstruction stability are reflected at the perceptual level. The visual assessment comprises the following dimensions:\u003c/p\u003e \u003cp\u003e \u003cem\u003eEdge continuity.\u003c/em\u003e Whether carved edges are continuous and intact, and whether the directional flow of wood grain appears smooth and coherent.\u003c/p\u003e \u003cp\u003e \u003cem\u003eDetail representation.\u003c/em\u003e Whether variations in carving depth are rendered naturally and whether fine-scale surface relief is adequately preserved.\u003c/p\u003e \u003cp\u003e \u003cem\u003eSurface roughness and noise aggregation.\u003c/em\u003e Whether high-frequency noise introduced during point cloud fusion, often resulting from accumulated redundant observations, leads to localized surface roughness or clustering of noise artifacts.\u003c/p\u003e \u003cp\u003eThe visual assessment used a five-point anchored rating scale (1 to 5), with each level corresponding to explicit quality criteria. A score of 1 indicates severe defects, making the model unsuitable for analysis or presentation. A score of 2 means only local features are discernible and overall quality is insufficient for practical use. A score of 3 indicates basic recognizability but evident fragmentation. A score of 4 represents a generally usable model with continuous major textures and geometric details, and minor defects that do not affect application. A score of 5 denotes excellent model quality with clear textures and well-preserved details. The final visual score was calculated as the average of the three visual dimensions. To ensure consistency, scores were averaged across ratings from five professionals in architectural heritage conservation (each with over 3 years of experience), with outlier ratings\u0026mdash;defined as those deviating significantly from the group mean\u0026mdash;excluded to mitigate subjective bias. The scoring examples are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3. Experiment and Application: architectural heritage component 3D reconstruction","content":"\u003cp\u003eTo validate the feasibility of the proposed lightweight, low-cost 3D modelling workflow, this study conducted experiments on eight types of architectural components differing in size, texture strength, geometric configuration, and colour richness. All experiments were carried out using standard consumer-grade hardware, without professional photography equipment or high-performance computing platforms, and all software tools used in the workflow were freely available.\u003c/p\u003e \u003cp\u003eAll experiments were performed on a MECHREVO Aurora X (standard edition) laptop, featuring a 13th-generation Intel Core i7 processor, an NVIDIA GeForce RTX 5060 Laptop GPU with 8 GB of dedicated memory, and 16 GB of system RAM.\u003c/p\u003e \u003cp\u003eImages were captured using a Redmi K80 smartphone with a 50 MP primary camera. All photos were taken in 1\u0026times; mode, with an equivalent focal length of 24 mm and a resolution of 4096 \u0026times; 3072 pixels. To ensure consistency, the focal length was fixed, exposure and white balance were locked, and HDR was disabled throughout image acquisition.\u003c/p\u003e \u003cp\u003ePoint cloud reconstruction was conducted with COLMAP version 3.12.6 (commit 4d5b60e), and mesh reconstruction as well as texture mapping were completed using MeshLab version 2025.07.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Components visualization and comparison\u003c/h2\u003e \u003cp\u003eTo further validate the applicability of the proposed modelling workflow across diverse component types and conditions, experiments were conducted on eight representative categories of architectural components with varying scales, texture characteristics, and geometric complexities. The selected components encompass typical scenarios such as low-texture, small-scale elements; high-texture, large-scale elements; and components with intricate geometric features, as summarized in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eOverview of basic characteristics, validation objectives, and model information of the reconstructed architectural components.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eID\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eType\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGeometry\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eValidation Objective\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eTime(min)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eSize(MB)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eResult\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTibetan tower exterior wall\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLarge-scale, low-texture applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e196\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTibetan-style canvas\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSmall-scale, high-colour applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e9.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e140\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eScreen wall base\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMedium\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLarge-scale, high-colour applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e218\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e617\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWooden doors and windows\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eMedium\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSmall-scale, wooden component applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e109\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWooden component\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLarge-scale, wooden component applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e182\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2070\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStone carving component\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSmall-scale, stone carving applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eStone carving component\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLarge-scale, stone carving applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eArchitectural balcony\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eLow\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eSimple geometric structure applicability\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e128\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe modelling results demonstrate that the proposed workflow consistently reproduces the principal geometric forms and texture features of architectural components across various types and characteristic conditions. As shown in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e, the resulting models meet the requirements for component-level digital documentation under different component scenarios. Figure\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents a comparison between the corresponding on-site photographs and the reconstructed three-dimensional models.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Threshold determination\u003c/h2\u003e \u003cp\u003eIn this experiment, the upper and lower modelling thresholds for each architectural component were established by analyzing the relationship between the number of input images and the corresponding Mean Track Length values during reconstruction. The upper threshold indicates the point at which models display clear, complete texture representation and well-defined geometric structures, while the lower threshold represents the minimum conditions required to preserve basic texture appearance and geometric integrity.\u003c/p\u003e \u003cp\u003eFrom a photogrammetric perspective, Mean Track Length characterizes the average co-visibility of three-dimensional points across multiple views, serving as a comprehensive indicator of observational redundancy, geometric constraint stability, and image network connectivity. A low Mean Track Length means that 3D points are supported by only a few viewing rays, rendering the bundle adjustment problem ill-posed and highly susceptible to outliers[19]. As Mean Track Length increases, multi-view redundancy is enhanced, significantly improving the stability of the geometric solution during bundle adjustment. However, beyond a certain threshold, the marginal utility of additional observations declines (i.e., diminishing returns). Excessive redundancy fails to yield geometric improvements and instead introduces high-frequency noise and matching ambiguities. Therefore, Mean Track Length theoretically has a reasonable effective range, providing a photogrammetric basis for threshold-based optimization of image acquisition strategies.\u003c/p\u003e \u003cp\u003eA total of 66 modelling experiments were conducted on various types of architectural components. The corresponding experimental data are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, and representative modelling results under different threshold conditions are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eEvaluation of modelling performance and threshold determination across MTL ranges for Components 1\u0026ndash;8\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"9\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eID\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMTL range\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDetail representation\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEdge continuity\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSurface roughness and noise\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eOverall visual score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eQuality assessment\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eUpper threshold\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c9\"\u003e \u003cp\u003eLower threshold\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncomplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscontinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNot usable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.84\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.84\u0026ndash;4.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.0\u0026ndash;5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;4.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncomplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscontinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNot usable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.46\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.46\u0026ndash;4.29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.0\u0026ndash;5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;4.29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncomplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscontinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNot usable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.59\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.59\u0026ndash;4.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.0\u0026ndash;5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;4.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncomplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscontinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNot usable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.71\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.71\u0026ndash;4.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.0\u0026ndash;5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;4.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncomplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscontinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNot usable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.75\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.75\u0026ndash;4.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.0\u0026ndash;5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;4.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncomplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscontinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNot usable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.54\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.54\u0026ndash;4.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.0\u0026ndash;5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;4.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC7\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncomplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscontinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNot usable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.63\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.63\u0026ndash;4.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.0\u0026ndash;5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;4.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eC8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;3.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eIncomplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eDiscontinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;4.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNot usable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c9\"\u003e \u003cp\u003e3.67\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3.67\u0026ndash;4.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eAbsent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.0\u0026ndash;5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eUsable\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e\u0026gt;\u0026thinsp;4.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eComplete\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContinuous\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePresent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSaturated\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c9\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAs shown in the figure, the upper threshold results displayed in the first and third rows preserve overall geometric integrity but increasingly exhibit redundant noise. In contrast, the lower threshold results in the second and fourth rows are characterized by geometric instability or loss of fine details.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Mean Track Length (MTL) calibration\u003c/h2\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e, despite clear differences among architectural components in scale, texture richness, and geometric complexity, the corresponding lower and upper threshold values exhibit strong clustering in their numerical distributions, with only minor fluctuations. Specifically, the lower thresholds of the tested components are mainly distributed within the range of approximately 3.5\u0026ndash;3.8, while the upper thresholds are concentrated around 4.1\u0026ndash;4.3, without significant dispersion or outliers. Based on these distribution patterns, the experimental results adopt 3.60\u0026ndash;4.20 as a representative Mean Track Length interval for component-level modelling, summarizing the clustered range observed across different component experiments.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eIn SfM-based image-based 3D reconstruction research, various metrics are available for evaluating reconstruction quality, such as Mean Track Length, point cloud density, reprojection error, and subjective visual assessment[20]. However, for component-level modelling of traditional architecture, most of these indicators serve as post-hoc evaluation results and cannot directly reflect the stability of multi-view geometric constraints or the impact of image redundancy on the reconstruction process. In this study, 66 practical modelling experiments were conducted on eight traditional architectural components with varying sizes, texture characteristics, and geometric complexities. The corresponding sparse reconstruction statistics for each experiment were systematically extracted.\u003c/p\u003e \u003cp\u003eUsing Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e as an example, the mean reprojection error remains consistently low as the number of images for each component increases. This is primarily because reprojection error serves as an optimization target during bundle adjustment. The solver minimizes residuals even when the underlying geometry is weak or ill-posed, provided that the feature correspondences are mathematically consistent. Consequently, low reprojection error does not necessarily guarantee structural stability.\u003c/p\u003e \u003cp\u003eFurther analysis reveals that although the Number of Observations rises with the number of images, its absolute values differ greatly among components due to variations in scale, texture, and coverage, making it unsuitable as a unified criterion for assessing modelling quality. In contrast, Mean Track Length offers a more consistent measure of the reliability of multi-view geometric constraints and exhibits stable variation patterns across different components. When Mean Track Length is low, reconstructed models frequently suffer from insufficient geometric constraints, leading to local structural instability, geometric collapse, and texture bleeding. Conversely, when Mean Track Length surpasses a certain range, even as the number of observations continues to rise, improvements in model quality tend to plateau, and excessive redundancy may introduce increased noise and surface roughness.\u003c/p\u003e \u003cp\u003eNotably, cross-component comparisons show that Mean Track Length exhibits strong correlation and consistency across various components, further confirming its value as a process-oriented indicator for component-level image-based 3D modelling.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBuilding upon the above analysis of indicator rationality, this study further defines an operational threshold range for Mean Track Length, which provides important engineering value for guiding image acquisition density control and reconstruction termination decisions. Within the COLMAP/SfM-based image-based 3D reconstruction framework, there has historically been a lack of stable intermediate indicators to guide image acquisition density and termination decisions during modelling. Most existing studies rely on point cloud size, reprojection error, or subjective visual assessment\u0026mdash;metrics that are primarily post hoc and offer limited process-level guidance. Given the established validity of Mean Track Length as a core indicator for component-level modelling, further defining its upper and lower thresholds is of significant engineering importance for optimizing image acquisition strategies and the reconstruction workflow.\u003c/p\u003e \u003cp\u003eGrounded in the stability of multi-view geometric constraints, this study introduces Mean Track Length as the core evaluation indicator for component-level modelling quality[21]. Drawing on 66 modelling experiments and analysis of 462 datasets, an operational Mean Track Length threshold range (3.60\u0026ndash;4.20) is proposed for traditional architectural components. Results show that when Mean Track Length drops below 3.60, stable geometric constraints are difficult to achieve, and the resulting models frequently present issues such as texture bleeding and local geometric collapse. When Mean Track Length exceeds 4.20, further increases in image redundancy fail to improve reconstruction quality and instead introduce redundant effects, including noise accumulation, point cloud redundancy, and increased reprojection error. Therefore, a Mean Track Length range of 3.60\u0026ndash;4.20 can be considered an optimal interval that balances modelling quality, computational efficiency, and cost control.\u003c/p\u003e \u003cp\u003eThis study demonstrates that, in component-level image-based modelling, blindly increasing image redundancy does not lead to a linear improvement in reconstruction reliability; instead, it may introduce unnecessary computational burdens and decision-making risks. This threshold transforms what was once an experience-dependent modelling workflow into a threshold-driven, quantifiable, and reproducible technical process. It offers clear criteria for image acquisition strategies and modelling decisions in the reconstruction of traditional architectural components, shifting the process from empirical methods to a controllable, threshold-based workflow. Consequently, it establishes a process control framework for component-level image-based modelling of architectural heritage that is grounded in explicit physical meaning and practical engineering applicability.\u003c/p\u003e \u003cp\u003eFrom a methodological perspective, the threshold-driven modeling strategy proposed in this study differs in its technical paradigm from the encapsulated photogrammetric solutions represented by commercial software. Existing commercial solutions are generally designed with encapsulated workflows to prioritize automation and user-friendliness. While efficient, this approach limits user intervention. In contrast, our proposed method offers process-level transparency, allowing for the explicit monitoring of geometric stability that is critical for scientific documentation.\u003c/p\u003e \u003cp\u003eIn contrast, this study develops a component-level modelling workflow using the open-source SfM framework COLMAP, with a core advantage in the interpretability and controllability of the reconstruction process. By introducing intermediate indicators such as Mean Track Length, the workflow enables continuous monitoring of feature matching performance and 3D geometric stability during reconstruction. This allows for real-time assessment of whether current image acquisition is within the \u0026ldquo;effective information range\u0026rdquo; and supports data-driven decisions on continuing or terminating image acquisition. The threshold-based (3.60\u0026ndash;4.20) process control mechanism transforms modelling failures from opaque outcomes into quantifiable, analysable, and correctable process-level issues.\u003c/p\u003e \u003cp\u003eIn summary, commercial modelling software and the approach proposed in this study represent two distinct technical paradigms, each tailored to different objectives rather than a straightforward comparison of superiority. Commercial solutions prioritize automation and result-oriented efficiency, while the proposed approach emphasizes process transparency, interpretability, and methodological reproducibility. For applications such as traditional architectural documentation, heritage surveys, and academic research, the threshold-driven modelling method introduced in this study offers a more controllable and verifiable technical alternative for component-level reconstruction.\u003c/p\u003e \u003cp\u003eDuring the analysis of the experimental data, a rather unexpected phenomenon was also observed. Existing studies generally suggest that texture-sparse or low-texture image regions increase the difficulty of local feature-based matching, thereby adversely affecting the quality of SfM reconstruction, as the number of extractable features and the reliability of feature matching are typically lower than those of high-texture components[22]. However, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e, no significant difference is observed in the lower thresholds between low-texture (C1) and high-texture (C6) components.\u003c/p\u003e \u003cp\u003eThis phenomenon may be attributed to the combined effects of multiple filtering processes and geometric compensation mechanisms during sparse reconstruction. On the one hand, in low-texture regions, although the number of extractable feature points is limited, those that remain after geometric consistency verification and bundle adjustment are typically more stable structural features, which tend to exhibit relatively longer track lengths, thereby statistically elevating the Mean Track Length of low-texture components[23]. On the other hand, some low-texture components simultaneously possess well-defined geometric structures, such as edges, corners, or component contours. These geometric constraints can partially compensate for the lack of texture information, making the effectiveness of multi-view geometric constraints comparable to that of high-texture components[24].\u003c/p\u003e \u003cp\u003eTo address the difficulty of point cloud extraction under low-texture conditions, this study further explored improvement strategies such as masking and secondary re-injection. However, these methods were only investigated as exploratory attempts and were not incorporated into the threshold-driven modelling framework proposed in this work.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAlthough this study proposes a Mean Track Length\u0026ndash;based threshold range for component-level image-based 3D modelling based on multiple component experiments, several limitations remain. First, the experimental samples are primarily drawn from eight representative architectural component types. While these samples offer diversity in texture richness, geometric complexity, and scale, the overall sample size is still relatively limited. Second, the proposed threshold framework is mainly designed for component-level image-based 3D modelling; its applicability to large-scale building or block-level modelling tasks requires further validation in future research. Additionally, as a statistical indicator derived from the sparse reconstruction stage, Mean Track Length (MTL) can be influenced by factors such as image resolution, feature extraction parameters, and image acquisition trajectories. Although this study maintained consistent parameter settings and modelling procedures throughout the experiments, the proposed threshold range may still vary under different hardware conditions or reconstruction configurations. Therefore, in practical applications, limited preliminary testing and calibration are recommended to adapt the threshold to specific workflows. Finally, the determination of the thresholds was based on a combined consideration of statistical indicators obtained from the sparse reconstruction stage and the actual modelling performance. However, the evaluation of model quality still lacks no-reference image quality assessment (No-Reference IQA) metrics from the perspective of image signal processing (ISP) to enable the automation of this process. The introduction and development of such metrics will therefore constitute an important direction for future work. Nevertheless, the judgment process in this study is constrained by the overall trends observed across multiple experimental results, thereby ensuring the rationality and interpretability of the derived thresholds within the scope of this research.\u003c/p\u003e"},{"header":"5. Conclusions and Prospects","content":"\u003cp\u003eThis study proposes a transferable, stability-threshold-centered decision-making approach for component-level image-based modelling. Designed to be low-cost, reproducible, and stable on standard computing devices, the workflow is intended for scenarios such as digital documentation of traditional architectural components and grassroots heritage surveys. Its engineering feasibility and process stability are demonstrated through experimental validation. Building on this workflow, the study introduces Mean Track Length as a threshold indicator for component-level image-based modelling and, through multiple experimental validations, establishes an operational effective range (3.60\u0026ndash;4.20). By establishing explicit, quantifiable criteria for acquisition density, this framework transforms component-level photogrammetry from an empirical, experience-dependent practice into a controllable, engineering-grade workflow.\u003c/p\u003e \u003cp\u003eFuture research can be pursued in several directions. Algorithmically, deeper exploration of the relationship between Mean Track Length and the internal multi-view geometric constraint mechanisms of SfM could provide a theoretical foundation for the threshold values proposed. In terms of application, the threshold-driven modelling strategy developed for component-level reconstruction in this study could be extended to broader heritage digitization scenarios, such as building fa\u0026ccedil;ades, individual structures, and even urban block\u0026ndash;scale contexts. At the indicator system level, Mean Track Length may be integrated with additional metrics, such as texture strength and viewpoint distribution, to establish a more comprehensive evaluation and decision-making system for component-level image-based 3D modelling quality. This would offer robust technical support for the digitalization of architectural heritage at larger scales and under more complex conditions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis work is supported by China-Portugal Joint Laboratory of Cultural Heritage Conservation Science\u0026nbsp;(No. SDYY2405),\u0026nbsp;Sichuan Science and Technology Program (Grant Number: 2025ZNSFSC1309),\u0026nbsp;and the Fundamental Research Funds for the Central Universities, Southwest Minzu University (No. 2024SYJSCX147, No. ZYN2025069).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDeclaration of Conflicting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no potential conflicts of interest concerning the research, authorship, and/or publication of this article.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor Contributions Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eLW conceived the research idea, designed the overall methodology, conducted the photogrammetric experiments, performed data analysis, and drafted the original manuscript; \u0026nbsp;YD contributed to data acquisition and image preprocessing, and assisted with experimental implementation and result validation; JJ participated in the development of the modelling workflow and supported data organization and visualization; YL assisted with experimental data collection and contributed to the evaluation of reconstruction results; HZ provided support in software operation, parameter testing, and preliminary result inspection; QY supervised and coordinated the research and reviewed and revised the manuscript; YZ provided academic guidance on research design, contributed to methodological refinement, and reviewed the manuscript.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eAll authors reviewed the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eLiu S, Bin Mamat MJ. Application of 3D laser scanning technology for mapping and accuracy assessment of the point cloud model for the Great Achievement Palace heritage building. Herit Sci. 2024;12(1):153.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMa H, Zhou Z, Wang Y. Digital preservation and development of architectural heritage from a virtual perspective: a systematic review. npj Herit Sci. 2025;13(1):594.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBuragohain D, Meng Y, Deng C, Li Q, Chaudhary S. Digitalizing cultural heritage through metaverse applications: challenges, opportunities, and strategies. Herit Sci. 2024;12(1):295.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng H, Chen L, Hu H, Wang Y, Wei Y. Research on the Digital Preservation of Architectural Heritage Based on Virtual Reality Technology. Buildings. 2024;14(5):1436.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGarc\u0026iacute;a-Molina DF, Gonz\u0026aacute;lez-Merino R, Rodero-P\u0026eacute;rez J, Carrasco-Hurtado B. 3D documentation for the conservation of historical heritage: the Castle of Priego de C\u0026oacute;rdoba (Spain). Virtual Archaeol Rev. 2021;12(24):115\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHe J, Jia Q. Opti3D for low light enhancement and calibration free 3D digitization of cultural relics. npj Herit Sci. 2025;13(1):357.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStylianidis E, Evangelidis K, Vital R, Dafiotis P, Sylaiou S. 3D Documentation and visualization of cultural heritage buildings through the application of geospatial technologies. Heritage. 2022;5(4):2818\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu J, Li B. Terrestrial laser scanning (TLS) survey and building information modeling (BIM) of the Edmund Pettus Bridge: A case study. Int Archives Photogrammetry Remote Sens Spat Inform Sci. 2024;48:379\u0026ndash;86.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRahaman H, Champion E. To 3D or not 3D: Choosing a photogrammetry workflow for cultural heritage groups. Heritage. 2019;2(3):1835\u0026ndash;51.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMarčiš M, Fraštia M, Terao Voškov\u0026aacute; K. Potential of low-cost UAV photogrammetry for documenting hard-to-access interior spaces through building openings. Heritage. 2024;7(11):6173\u0026ndash;91.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNeumann KA, Tausch R, Kutlu H, Kuijper A, Santos P, Fellner D. Point cloud quality metrics for incremental image-based 3D reconstruction. Multimedia Tools Appl. 2025;84(32):39123\u0026ndash;41.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCalisi D, Botta S, Cannata A. Integrated surveying, from laser scanning to UAV systems, for detailed Documentation of architectural and archeological heritage. Drones. 2023;7(9):568.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKlapa P, Żygadło A, Pepe M. 3D heritage reconstruction through HBIM and multi-source data fusion: geometric change analysis across decades. Appl Sci. 2025;15(16):8929.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMaiwald F, Feurer D, Eltner A. Solving photogrammetric cold cases using AI-based image matching: New potential for monitoring the past with historical aerial images. ISPRS J Photogrammetry Remote Sens. 2023;206:184\u0026ndash;200.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSchonberger JL, Frahm J-M. Structure-from-motion revisited. In: \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition\u003c/em\u003e: 2016; 2016: 4104\u0026ndash;4113.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCignoni P, Callieri M, Corsini M, Dellepiane M, Ganovelli F, Ranzuglia G. Meshlab: an open-source mesh processing tool. In: \u003cem\u003eEurographics Italian chapter conference: 2008\u003c/em\u003e: Salerno; 2008: 129\u0026ndash;136.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePataki Z, Sarlin P-E, Sch\u0026ouml;nberger JL, Pollefeys M. MP-SfM: Monocular Surface Priors for Robust Structure-from-Motion. In: \u003cem\u003eProceedings of the Computer Vision and Pattern Recognition Conference\u003c/em\u003e: 2025; 2025: 21891\u0026ndash;21901.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHeim RH, Okole N, Steppe K, Van Labeke M-C, Geedicke I, Maes WH. An applied framework to unlocking multi-angular UAV reflectance data: a case study for classification of plant parameters in maize (Zea mays). Precision Agric. 2024;25(3):1751\u0026ndash;75.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCueto Zumaya CR, Catalano I, Queralta JP. Building Better Models: Benchmarking Feature Extraction and Matching for Structure from Motion at Construction Sites. Remote Sens. 2024;16(16):2974.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHou Q, Xia R, Zhang J, Feng Y, Zhan Z, Wang X. Learning visual overlapping image pairs for SfM via CNN fine-tuning with photogrammetric geometry information. Int J Appl Earth Obs Geoinf. 2023;116:103162.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang D, Qin R, Elhashash M. Bundle adjustment with motion constraints for uncalibrated multi-camera systems at the ground level. ISPRS J Photogrammetry Remote Sens. 2024;211:452\u0026ndash;64.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang X, Jiang G. A practical 3D reconstruction method for weak texture scenes. Remote Sens. 2021;13(16):3103.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee J, Yoo S. Dense-SfM: Structure from Motion with Dense Consistent Matching. In: \u003cem\u003eProceedings of the Computer Vision and Pattern Recognition Conference\u003c/em\u003e: 2025; 2025: 6404\u0026ndash;6414.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao L, Guo F, Zhu Y, Wang H, Zhou B. A Generalized Voronoi Diagram-Based Segment-Point Cyclic Line Segment Matching Method for Stereo Satellite Images. Remote Sens. 2024;16(23):4395.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"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":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Architectural heritage, Photogrammetry, Mean Track Length, Quantification threshold framework, COLMAP","lastPublishedDoi":"10.21203/rs.3.rs-8709145/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8709145/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eExisting photogrammetric workflows for architectural heritage documentation often lack process-oriented quantitative criteria to guide image acquisition, resulting in reconstruction failure due to data sparsity or noise and inefficiency caused by excessive image acquisition. This study proposes a quantitative threshold framework based on Mean Track Length (MTL) for component-level 3D reconstruction using COLMAP. A total of 66 reconstruction experiments and 462 data records were conducted on eight types of traditional architectural components. The results indicate that reliable reconstruction is achieved when MTL lies within 3.60\u0026ndash;4.20; values below this range lead to unstable geometry, whereas values above it do not improve reconstruction quality and instead introduce noise. This study establishes a measurable stopping criterion for image acquisition and transforms reconstruction from an experience-driven practice into a threshold-driven and controllable workflow, enabling reproducible and efficient architectural heritage documentation.\u003c/p\u003e","manuscriptTitle":"Novel Mean Track Length-Driven 3D Framework via COLMAP for Architectural Heritage Photogrammetry","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-12 11:05:33","doi":"10.21203/rs.3.rs-8709145/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-07T17:23:47+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-05T11:54:41+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-02-24T13:03:15+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"20163660389867343661225616684017791361","date":"2026-02-09T07:49:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"183812754776208939224099017359501632896","date":"2026-02-08T16:08:38+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-08T14:44:52+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-02-03T22:38:19+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-02-02T04:16:07+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Heritage Science","date":"2026-01-27T09:56:36+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-heritage-science","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"hsci","sideBox":"Learn more about [Heritage Science](http://heritagesciencejournal.springeropen.com)","snPcode":"40494","submissionUrl":"https://submission.nature.com/new-submission/40494/3","title":"npj Heritage Science","twitterHandle":"@SpringerOpen","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"54f6a6dc-ffcf-4d72-a623-0a5226d3b8cb","owner":[],"postedDate":"February 12th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-12T18:30:42+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-12 11:05:33","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8709145","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8709145","identity":"rs-8709145","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}