Improvement of 3D reconstruction based on a new 3D point cloud filtering algorithm

Abstract

Structure from motion (SfM) is a 3D reconstruction approach to recover a camera pose and 3D coordinates of matched interest points. The obtained 3D structure is not clear. We must therefore use methods that significantly increase the number of reconstructed 3D points. Among these methods are those based on the match propagation. However, the 3D point recovery process generates erroneous points due to false matches. In this work, we propose a new algorithm to eliminate these wrong reconstructed 3D points. Our algorithm allows to improve the quality of the 3D reconstruction in lower calculation time. At first, the SfM approach is used to recover the sparse 3D structure. Afterward, we apply the Modified Match Propagation algorithm on image couples to retrieve new matches and their 3D coordinates. The matching result is used to define the 3D point neighborhoods. These neighborhoods, the barycenter and the Euclidean distance will be used to eliminate the erroneous 3D points. The final 3D model can be obtained with meshing and texture mapping. Experimental results show the efficiency and the rapidity of the proposed approach.