A no-reference method for face-level defect detection and quality assessment of oblique photogrammetric 3D mesh models

doi:10.11947/j.AGCS.2026.20250517

Abstract

Abstract:

Addressing the urgent need for quality control of large-scale oblique photogrammetric models in the real-scene 3D China initiative, this study focuses on two limitations of existing methods: full-reference methods depend on costly reference data, whereas conventional no-reference methods generally lack the ability to detect defects at the individual mesh-face level. We propose a no-reference method for face-level defect detection and quality assessment of oblique photogrammetric 3D mesh models based on intrinsic consistency and a graph attention network. Although the method requires no external reference model during inference, its detection performance is evaluated against face-level ground-truth labels. Based on the imaging characteristics of oblique photogrammetry and the topological constraints of 3D meshes, we construct a set of multimodal features describing photometric consistency, multi-view geometric structure, triangle visibility, local point-cloud density, normal consistency, and triangle shape. The mesh is then represented as a graph, and a graph attention mechanism aggregates contextual information from neighboring faces to support adaptive inference of local mesh quality. Three representative regional datasets containing different types of surface features were used to evaluate detection accuracy. Generalization and computational efficiency were further evaluated on large-scale oblique photogrammetric mesh models. The proposed method effectively detects typical defects, including geometric noise, geometric deformation and stretching, and incomplete geometry. Across different scene types, it achieves a defect detection accuracy of 71%~78% and recall rates above 80%. For meshes containing millions of faces, both training and inference can be completed within minutes. The model also maintains reliable detection performance in previously unseen scenes. In addition, multi-level downsampling improves detection efficiency while preserving relatively high detection accuracy. These results provide a practical approach to automated quality inspection and fine-grained defect localization in large-scale 3D mesh data, supporting comprehensive quality assessment of real-scene 3D models.

Key words: real-scene 3D, oblique photogrammetric model, no-reference quality assessment, graph attention network, defect detection

CLC Number:

P208

Shengjun Tang, Hanyu Li, Weixi Wang, Linfu Xie, Xiaoming Li, Renzhong Guo. A no-reference method for face-level defect detection and quality assessment of oblique photogrammetric 3D mesh models[J]. Acta Geodaetica et Cartographica Sinica, 2026, 55(8): 1382-1399.

Add to citation manager EndNote|Reference Manager|ProCite|BibTeX|RefWorks

URL: http://xb.chinasmp.com/EN/10.11947/j.AGCS.2026.20250517

http://xb.chinasmp.com/EN/Y2026/V55/I8/1382

Figures/Tables 17

Fig. 1

Tab. 1

Fig. 2

Tab. 2

Fig. 3

Fig. 4

Tab. 3

Tab. 4

Fig. 5

Fig. 6

Fig. 7

Tab. 5

Fig. 8

Tab. 6

Fig. 9

Fig. 10

Tab. 7

References 32

[1]	杨必胜, 陈驰, 董震. 面向智能化测绘的城市地物三维提取[J]. 测绘学报, 2022, 51(7): 1476-1484. DOI: . doi: 10.11947/j.AGCS.2022.20220183
	Yang Bisheng, Chen Chi, Dong Zhen. 3D geospatial information extraction of urban objects for smart surveying and mapping[J]. Acta Geodaetica et Cartographica Sinica, 2022, 51(7): 1476-1484. DOI: . doi: 10.11947/j.AGCS.2022.20220183
[2]	陈军, 刘建军, 田海波. 实景三维中国建设的基本定位与技术路径[J]. 武汉大学学报(信息科学版), 2022, 47(10): 1568-1575.
	Chen Jun, Liu Jianjun, Tian Haibo. Basic directions and technological path for building 3D realistic geospatial scene in China[J]. Geomatics and Information Science of Wuhan University, 2022, 47(10): 1568-1575.
[3]	顾建祥, 董震, 郭王. 面向上海城市数字化转型的新型测绘[J]. 测绘通报, 2021(7): 131-134, 139.
	Gu Jianxiang, Dong Zhen, Guo Wang. New surveying and mapping for the urban digital transformation of Shanghai[J]. Bulletin of Surveying and Mapping, 2021(7): 131-134, 139.
[4]	肖建华, 李海亭, 李鹏鹏, 等. 实景三维的内涵与分类分级[J]. 城市勘测, 2021(5): 5-10.
	Xiao Jianhua, Li Haiting, Li Pengpeng, et al. Connotation, classification and grading of 3D real scene[J]. Urban Geotechnical Investigation & Surveying, 2021(5): 5-10.
[5]	朱庆, 朱军, 黄华平, 等. 实景三维空间信息平台与数字孪生川藏铁路[J]. 高速铁路技术, 2020, 11(2): 46-53.
	Zhu Qing, Zhu Jun, Huang Huaping, et al. Real scene 3D spatial information platform and digital twin Sichuan-Tibet railway[J]. High Speed Railway Technology, 2020, 11(2): 46-53.
[6]	李清泉, 邵成立, 万剑华, 等. 优视摄影测量与泛在实景三维数据采集：以实景三维青岛为例[J]. 武汉大学学报(信息科学版), 2022, 47(10): 1587-1597.
	Li Qingquan, Shao Chengli, Wan Jianhua, et al. Optimized views photogrammetry and ubiquitous real 3D data acquisition with the application case in Qingdao[J]. Geomatics and Information Science of Wuhan University, 2022, 47(10): 1587-1597.
[7]	肖雄武. 具备结构感知功能的倾斜摄影测量场景三维重建[J]. 测绘学报, 2019, 48(6): 802. DOI: . doi: 10.11947/j.AGCS.2019.20180319
	Xiao Xiongwu. Oblique photogrammetry based scene 3D reconstruction with structure sensing functions[J]. Acta Geodaetica et Cartographica Sinica, 2019, 48(6): 802. DOI: . doi: 10.11947/j.AGCS.2019.20180319
[8]	李德仁, 肖雄武, 郭丙轩, 等. 倾斜影像自动空三及其在城市真三维模型重建中的应用[J]. 武汉大学学报(信息科学版), 2016, 41(6): 711-721.
	Li Deren, Xiao Xiongwu, Guo Bingxuan, et al. Oblique image based automatic aerotriangulation and its application in 3D city model reconstruction[J]. Geomatics and Information Science of Wuhan University, 2016, 41(6): 711-721.
[9]	魏东, 刘欣怡, 张永军. 多视影像三维线云重建技术及其智能化发展展望[J]. 测绘学报, 2024, 53(6): 1025-1036. DOI: . doi: 10.11947/j.AGCS.2024.20230447
	Wei Dong, Liu Xinyi, Zhang Yongjun. The technology and intelligent development of 3D line cloud reconstruc-tion from multiple images[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(6): 1025-1036. DOI: . doi: 10.11947/j.AGCS.2024.20230447
[10]	肖勇, 王成, 习晓环, 等. 机载激光雷达数据的建筑物三维模型重建[J]. 测绘科学, 2014, 39(11): 37-41.
	Xiao Yong, Wang Cheng, Xi Xiaohuan, et al. 3D building model reconstruction from airborne LiDAR data[J]. Science of Surveying and Mapping, 2014, 39(11): 37-41.
[11]	刘欣怡, 张永军, 范伟伟, 等. 无人机倾斜摄影三维建模技术研究现状及展望[J]. 时空信息学报, 2023, 30(1): 41-48.
	Liu Xinyi, Zhang Yongjun, Fan Weiwei, et al. 3D modeling based on UAV oblique photogrammetry: research status and prospect[J]. Journal of Spatio-Temporal Information, 2023, 30(1): 41-48.
[12]	朱庆, 张利国, 丁雨淋, 等. 从实景三维建模到数字孪生建模[J]. 测绘学报, 2022, 51(6): 1040-1049. DOI: . doi: 10.11947/j.AGCS.2022.20210640
	Zhu Qing, Zhang Liguo, Ding Yulin, et al. From real 3D modeling to digital twin modeling[J]. Acta Geodaetica et Cartographica Sinica, 2022, 51(6): 1040-1049. DOI: . doi: 10.11947/j.AGCS.2022.20210640
[13]	单杰, 李志鑫, 张文元. 大规模三维城市建模进展[J]. 测绘学报, 2019, 48(12): 1523-1541. DOI: . doi: 10.11947/j.AGCS.2019.20190471
	Shan Jie, Li Zhixin, Zhang Wenyuan. Recent progress in large-scale 3D city modeling[J]. Acta Geodaetica et Cartographica Sinica, 2019, 48(12): 1523-1541. DOI: . doi: 10.11947/j.AGCS.2019.20190471
[14]	王丙涛, 王继. 基于倾斜摄影技术的三维建模生产与质量分析[J]. 城市勘测, 2015(5): 80-82, 85.
	Wang Bingtao, Wang Ji. Production and quality analysis of 3D city modeling using oblique photogrametric technology[J]. Urban Geotechnical Investigation & Surveying, 2015(5): 80-82, 85.
[15]	Verdie Y, Lafarge F, Alliez P. LOD generation for urban scenes[J]. ACM Transactions on Graphics, 2015, 34(3): 1-14.
[16]	李清泉, 黄惠, 姜三, 等. 优视摄影测量方法及精度分析[J]. 测绘学报, 2022, 51(6): 996-1007. DOI: . doi: 10.11947/j.AGCS.2022.20210685
	Li Qingquan, Huang Hui, Jiang San, et al. Optimized views photogrammetry and its precision analysis[J]. Acta Geodaetica et Cartographica Sinica, 2022, 51(6): 996-1007. DOI: . doi: 10.11947/j.AGCS.2022.20210685
[17]	赵建春. 基于倾斜摄影测量技术的实景三维建模及精度分析[J]. 科学技术创新, 2022(7): 127-130.
	Zhao Jianchun. 3D modeling and precision analysis of real scene based on oblique photogrammetry technology[J]. Scientific and Technological Innovation, 2022(7): 127-130.
[18]	Sorgente T, Biasotti S, Manzini G, et al. A survey of indicators for mesh quality assessment[J]. Computer Graphics Forum, 2023, 42(2): 461-483.
[19]	杨斌, 李晓强, 李伟, 等. 基于表面粗糙度的三维模型质量评价研究[J]. 计算机科学, 2011, 38(1): 276-278, 285.
	Yang Bin, Li Xiaoqiang, Li Wei, et al. Quality assessment for 3D model based on surface roughness[J]. Computer Science, 2011, 38(1): 276-278, 285.
[20]	张号, 王炜, 邓强, 等. 倾斜摄影实景三维模型质量评价[J]. 北京测绘, 2020, 34(1): 56-60.
	Zhang Hao, Wang Wei, Deng Qiang, et al. The quality evaluation of 3D model of tilt photography reality[J]. Beijing Surveying and Mapping, 2020, 34(1): 56-60.
[21]	张雯, 程亮, 张群, 等. 三维模型几何质量评价方法[J]. 测绘通报, 2014(7): 44-47.
	Zhang Wen, Cheng Liang, Zhang Qun, et al. An evaluation method for the geometry quality of three-dimensional models[J]. Bulletin of Surveying and Mapping, 2014(7): 44-47.
[22]	钟炜, 郭丽. 城市级实景三维单体化模型质检方法研究[J]. 地理空间信息, 2025, 23(7): 133-136.
	Zhong Wei, Guo Li. Research on quality inspection method for urban 3D real scene monolithic model[J]. Geospatial Information, 2025, 23(7): 133-136.
[23]	Maraş E E, Nasery N. Investigating the length, area and volume measurement accuracy of UAV-Based oblique photogrammetry models produced with and without ground control points[J]. International Journal of Engineering and Geosciences, 2023, 8(1): 32-51.
[24]	Xu Y M, Zhang J X, Zhao H T, et al. Research on quality framework of real scene 3D model based on oblique photogrammetry[J]. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, 2021, XLIII-B3-2021: 791-796.
[25]	Abouelaziz I, El Hassouni M, Cherifi H. No-reference 3D mesh quality assessment based on dihedral angles model and support vector regression[C]//Proceedings of 2016 Image and Signal Processing. Cham: Springer International Publishing, 2016: 369-377.
[26]	Nouri A, Charrier C, Lézoray O. 3D blind mesh quality assessment index[C]//Proceedings of 2017 IS&T International Symposium on Electronic Imaging. Burlingame: Society for Imaging Science and Technology, 2017: 9-26.
[27]	Nehmé Y, Delanoy J, Dupont F, et al. Textured mesh quality assessment: large-scale dataset and deep learning-based quality metric[J]. ACM Transactions on Graphics, 2023, 42(3): 1-20.
[28]	Zhang Zicheng, Sun Wei, Min Xiongkuo, et al. No-reference quality assessment for 3D colored point cloud and mesh models[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 32(11): 7618-7631.
[29]	Dong Lu, Fang Yuming, Lin Weisi, et al. Perceptual quality assessment for 3D triangle mesh based on curvature[J]. IEEE Transactions on Multimedia, 2015, 17(12): 2174-2184.
[30]	Ibork Z, Nouri A, Lézoray O, et al. No reference 3D mesh quality assessment using deep convolutional features[C]//Proceedings of 2023 International Symposium on Image and Signal Processing and Analysis. Rome: IEEE, 2023: 1-6.
[31]	Liu Qi, Yuan Hui, Su Honglei, et al. PQA-net: deep no reference point cloud quality assessment via multi-view projection[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2021, 31(12): 4645-4660.
[32]	Dong S, Yan Q, Xu Y, et al. Quality inspection and analysis of three-dimensional geographic information model based on oblique photogrammetry[C]//Proceedings of 2018 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Beijing: Copernicus Publications, 2018: 299-302.

特征名称	符号	数据源	物理含义	针对的主要缺陷类型
光度一致性	x_photo	影像+网格	多视纹理匹配的置信度	几何噪声、模糊、错位
多视几何结构	x_geo	影像+网格	空间前方交会的几何稳定性	深度误差大、几何噪点
三角形可见性	x_vis	影像+网格	参与重建的观测冗余度	边缘悬浮物、虚假几何
局部点云密度	x_den	网格	原始测量数据的支持程度	平面大孔洞填充、非真实表面
法向一致性	x_norm	网格	网格面与原始流形的拟合度	表面凹凸不平、拓扑翻转
三角形形状	x_shape	网格	网格单元的几何健康度	狭长三角形、退化网格

数据集名称	场景描述	待检测模型	真值模型	主要质量挑战
数据集1：ISPRS多特蒙德数据集	包含同步采集的机载倾斜与地面影像	机载影像重建	空地融合高质量Mesh模型	待检测模型在建筑立面、底部及遮挡区域存在明显的结构缺失和纹理变形
数据集2：复杂单体建筑数据集	具有显著特殊几何特征的单体建筑	机载影像重建	人工精细化建模得到的高精度模型	考验方法针对特定复杂几何结构的检测能力，模型易出现几何粘连或结构错误
数据集3：复杂建筑群数据集	大范围、高密度的城市建筑群体场景	机载影像重建	人工精细化建模完成的高精度模型	考察方法在大规模场景、严重相互遮挡条件下，对多样化建筑形态质量问题的稳健性

数据集名称	总网格数	正样本网格数	负样本网格数	训练集总数	训练集占比
数据集1	1 242 784	70 330	2 874	73 204	1/17
数据集2	944 524	145 600	5 775	151 375	1/6
数据集3	6 642 898	496 822	15 076	511 898	1/13

试验方法	数据集场景	数据规模		检测精度指标				效率指标
试验方法	数据集场景	测试网格数	辅助影像数	Precision/（%）	Recall/（%）	F₁值/（%）	IoU/（%）	训练时间/s	推理时间/s
Mesh-GAT	数据集1	1 242 784	444	78.1	84.5	81.2	68.3	40	50
	数据集2	944 524	337	71.5	81.2	76.0	61.3	23	40
	数据集3	6 642 898	235	76.7	89.4	82.6	70.3	130	319
随机森林	数据集1	1 242 784	444	73.3	76.7	75.0	60.0	3	2
	数据集2	944 524	337	63.4	72.1	67.5	50.9	2	2
	数据集3	6 642 898	235	64.2	78.5	70.6	54.6	12	6

特征设置	Precision	Recall	F₁值	IoU
完整六维特征	76.7	89.4	82.6	70.3
去除三角形可见性	72.4	78.6	75.4	60.5
去除多视几何结构	63.8	91.2	75.1	60.1
去除光度一致性	68.5	73.1	70.7	54.7
去除三角顶点局部密度	74.2	68.4	71.2	55.3
去除三角形与点云法向一致性	64.9	59.8	62.2	45.2
去除三角形形状	70.6	70.2	70.4	54.3
仅使用视觉特征	80.3	61.5	69.7	53.4
仅使用几何特征	82.1	53.6	64.9	48.0