A Practical 3D Reconstruction Method for Weak Texture Scenes

Xuyuan Yang,Guang Jiang

doi:10.3390/rs13163103

Abstract

In recent years, there has been a growing demand for 3D reconstructions of tunnel pits, underground pipe networks, and building interiors. For such scenarios, weak textures, repeated textures, or even no textures are common. To reconstruct these scenes, we propose covering the lighting sources with films of spark patterns to “add” textures to the scenes. We use a calibrated camera to take pictures from multiple views and then utilize structure from motion (SFM) and multi-view stereo (MVS) algorithms to carry out a high-precision 3D reconstruction. To improve the effectiveness of our reconstruction, we combine deep learning algorithms with traditional methods to extract and match feature points. Our experiments have verified the feasibility and efficiency of the proposed method.

Highlights

With the increasing demand for measuring scenes like tunnel pits, underground pipe networks, and interior scenarios of buildings, more and more studies have focused on precisely measuring such structures
One of the most widely used techniques for precise reconstruction of scenes is a combination of structure from motion (SFM) [1] and multiview stereo (MVS) [2]
After images of the scenes are captured, the sparse point cloud and camera poses can be calculated by the SFM algorithm

Summary

Introduction

With the increasing demand for measuring scenes like tunnel pits, underground pipe networks, and interior scenarios of buildings, more and more studies have focused on precisely measuring such structures. After images of the scenes are captured, the sparse point cloud and camera poses can be calculated by the SFM algorithm. After the SFM algorithm is completed, the resulting sparse point cloud and extrinsic camera parameters are taken as the MVS algorithm’s input. In the scenes with weak textures, repeated textures, or no textures (we use “weak texture” to refer to these situations), such as tunnel pits, underground pipe networks, and building interiors, the feature matching–based SFM algorithm shows poor results because traditional feature extraction algorithms cannot extract enough reliable feature points. Even if the SFM can obtain the correct camera pose through correctly matched feature pairs, most of the pixels in the images containing weak texture do not meet the constraint of color consistency.

Results

Discussion

Conclusion