Research on Optimization of Fine 3D Reconstruction Process Based on SfM-MVS

Abstract

Aiming at the optimization problem of motion recovery structure and multi-view stereo (MVS) vision process, this study systematically explores the influence of shooting parameters on 3D reconstruction quality by establishing data acquisition standards, filling the gap of standardization research in this field. This study aims to quantify the effect of variables such as lighting conditions, shooting equipment, and background complexity on reconstruction accuracy and propose a scientific acquisition standard. Methods: Four groups of control experiments (control group and three experimental groups) were designed. Based on Reality Capture, the reconstruction effects under ambient light/point light source, professional camera/mobile phone, solid color/complex background, and other conditions were compared and analyzed, and the success rate of alignment, point cloud density, and other indicators were evaluated. The experimental results show that the alignment success rate of uniform lighting group is 78.7% (only 12.6% for point light source), the point cloud density is reduced by 40% due to automatic exposure of mobile phone, the alignment rate of complex background group can be restored to 60.5% through control point optimization, and the number of surfaces in the reflective area is constant to 50% in the non-reflective area. A hybrid optimization scheme combining traditional preprocessing (such as High Dynamic Range Imaging (HDR) correction) and neural network (such as Deep Neural Network for Image Denoising (DnCNN)) is further proposed.