IMPROVED MULTISTAGE LEARNING FOR MULTIBODY MOTION SEGMENTATION

Abstract: We present an improved version of the MSL method of Sugaya and Kanatani for multibody motion segmen-tation. We replace their initial segmentation based on heuristic clustering by an analytical computation based on GPCA, fitting two 2-D affine spaces in 3-D by the Taubin method. This initial segmentation alone can segment most of the motions in natural scenes fairly correctly, and the result is successively optimized by the EM algorithm in 3-D, 5-D, and 7-D. Using simulated and real videos, we demonstrate that our method out-performs the previous MSL and other existing methods. We also illustrate its mechanism by our visualization technique