NON-rigid structure from motion via sparse self-expressive representation

Abstract

To simultaneously recover 3D shapes of non-rigid object and camera motions from 2D corresponding points is a difficult task in computer vision. This task is called Non-rigid Structure from motion(NRSfM). To solve this ill-posed problem, many existing methods rely on low rank assumption. However, the value of rank has to be accurately predefined because incorrect value can largely degrade the reconstruction performance. Unfortunately, these is no automatic solution to determine this value. In this paper, we present a self-expressive method that models 3D shapes with a sparse combination of other 3D shapes from the same structure. One of the biggest advantages is that it doesn't need the rank to be predefined. Also, unlike other learning-based methods, our method doesn't need learning step. Experimental results validate the efficiency of our method.