基于单幅图像学习景深的古建局部三维重建和精度评测

本文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，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，本文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二种途径，比较了在双目相机固定的图像获取方式下和同时估计相机运动的单相机图像获取方式下，基于单幅图像学习景深的精度差异。实验结果表明，尽管结构和重复纹理现象在基于多幅图像的三维重建中是一个困难的问题，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的重建精度。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 whether current unsupervised single-image depth estimation methods can effectively address the inherent structural and repetitive texture issues in ancient architecture images, and whether they can achieve the centimeter-level reconstruction accuracy required for ancient architecture archiving. Specifically, this paper takes data acquired by structured-light depth cameras as the ground truth, and compares the accuracy differences of single-image depth estimation methods under two image acquisition scenarios: fixed binocular camera setup, and single-camera image acquisition with simultaneous camera motion estimation, via two direct comparison approaches: depth maps and 3D point clouds. The experimental results show that although structural and repetitive texture issues are a challenging problem in multi-image-based 3D reconstruction, their impact on single-image depth estimation is generally not significant. Additionally, although single-image depth estimation methods have achieved accuracy comparable to laser scanning on many public indoor and outdoor datasets, they still struggle to meet the centimeter-level reconstruction accuracy required for digital archiving of ancient architecture in 3D reconstruction tasks. Subsequent research needs to further explore approaches to improve reconstruction accuracy, particularly methods based on model prior constraints.

搜集汇总

背景与挑战

背景概述

该数据集聚焦于从单幅图像学习景深以实现古建筑局部三维重建和精度评测的研究。研究评估了无监督学习方法在处理古建筑图像结构和纹理重复方面的能力，并探讨了其是否满足古建筑数字存档所需的厘米级精度要求。数据集包含通过结构光深度相机获取的真实数据，用于比较和评估不同场景下的深度学习精度。