基于单幅图像学习景深的古建局部三维重建和精度评测  

3D Reconstruction and Accuracy Evaluation of Ancient Chinese Architectural Patches Based on Depth Learning from Single Image

作　　者：胡立华[1,2] 阴文庄 邢思远 张继福[1] 董秋雷[2] 胡占义[2] Hu Lihua;Yin Wenzhuang;Xing Siyuan;Zhang Jifu;Dong Qiulei;Hu Zhanyi(College of Computer Sciences and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,Shanxi,China;National Laboratory of Pattern Recognition,Institute of Automation,Chinese Academy of Sciences,Beijing 100190,China)

机构地区：[1]太原科技大学计算机科学与技术学院,山西太原030024 [2]中国科学院自动化研究所模式识别国家重点实验室,北京100190

出　　处：《激光与光电子学进展》2022年第14期216-225,共10页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(U1805264)。

摘　　要：主要探索当前无监督框架下基于单幅图像学习景深的方法,研究该方法是否能有效应对古建图像固有的结构和纹理重复现象,以及能否达到古建存档要求的厘米级重建精度。具体地,将结构光深度相机获取的数据作为真值,通过直接比较深度图和三维点云,比较在双目相机固定的图像获取方式下和在单相机运动的图像获取方式下基于单幅图像学习景深的精度差异。实验结果表明,尽管因结构和纹理重复现象的存在基于多幅图像的三维重建是一个困难的问题,但对基于单幅图像学习景深的影响一般并不明显。另外,尽管基于单幅图像学习景深在很多公开的室内和室外数据集上均取得了与激光扫描相媲美的精度,但对古建三维重建而言,目前仍难以达到古建数字化存档要求的厘米级重建精度。后续需要进一步探索提高重建精度的途径,特别是基于模型先验约束的方法。This paper primarily explores the depth learning method from a single image under the current unsupervised framework. It investigates whether this method can effectively deal with the repetition of the inherent structure and texture of ancient Chinese architectural images and whether it can meet the centimeter-level reconstruction accuracy required by the Chinese architecture documentation standard. Specifically, the accuracy difference of depth learning based on a single image under the image acquisition mode of fixed binocular cameras and the image acquisition mode of a single moving camera is compared using the data obtained by the structured light depth camera as the ground truth by directly comparing the depth map and the three-dimensional(3D) point cloud. The experimental results show that while 3D reconstruction based on multiple images is challenging due to the existence of repeated structures and textures, the impact of the existence on depth learning based on a single image is generally insignificant. In addition, even though depth learning based on a single image has achieved comparable accuracy with laser scanning on many open indoor and outdoor datasets, it is still difficult to achieve the centimeter-level reconstruction accuracy required by the digital documentation standard of ancient Chinese architectural3D reconstruction. In the future, the shape of prior information will be exploited to improve the reconstruction accuracy.

关 键 词：机器视觉 古建 景深学习 精度评估 三维重建 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]