Comparative Study of Structure-from-Motion Algorithms 3D Shape Reconstruction

Abstract

3D shape reconstruction from multiview photographs and video sequences (2D images) is an active area of research. Existing face recognition systems are based on 2D facial images and exhibit well-known deficiencies. Accordingly, the face recognition research is gradually shifting from classical 2D to sophisticated 3D or hybrid 2D/3D. Currently the 3D reconstruction algorithms may be grouped in to four categories. These are shape-from-X, 3D morphable model (3DMM), structure from motion (SFM) and learning based. In this paper, we introduce, discuss and analyze the recent SFM methods, depth estimation based on genetic algorithm (SM), constrained independent component analysis (cICA) and non-linear least squares modeling (NLS). We begin by introducing similarity transform, which forms the basis of SFM reconstruction techniques described here. This is followed by a review and comparison of the three methods. The characteristics of the three SFM methods are summarized in a table that should facilitate further research on this topic.