Multi-body segmentation and motion number estimation via over-segmentation detection

Abstract

This paper studies the problem of multi-body segmentation and motion number estimation. It is well known that motion number plays a critical role in the success of multi-body segmentation. Most of the existing methods exploit only motion affinity to segment and determine the number of motions. Motion number estimated in this way is often seriously affected by noise. In this paper, we recast the problem of multi-body segmentation and motion number estimation into an over-segmentation detection problem, and introduce three measures, namely loss of spatial locality (LSL), split ratio (SR) and cluster distance (CD), for over-segmentation detection. A hierarchical clustering method based on motion affinity is applied to split the motion clusters recursively until over-segmentation occurs. Over-segmentation is detected by Kernel Support Vector Machines trained under supervised learning using the above three measures. We leverage on Hopkins155 database to test our method and, with the same motion affinity measure, our method outperforms another state-of-the-art method. To the best of our knowledge, this paper is the first to tackle the problem of multi-body segmentation and motion number estimation from the perspective of over-segmentation detection. © 2011 Springer-Verlag Berlin Heidelberg.