Automatic Segmentation of Articulated Objects of Unknown Complexity from Videos

Abstract

Motion segmentation is an important topic in computer vision. Unlike sim- ple motion segmentation articulated segmentation gives much more information about observed scene and makes easier classification of the observed object, i.e. by kinetic patterns. We are interested in automatic segmentation of an artic- ulated object with an unknown complexity. Typical representative is a human being or a more complicated machine. We implemented the method described in M. Pawan Kumar, P.H.S. Torr a A. Zisserman - Learning Layered Motion Segmentation of Video. This algorithm is applicable to any video containing piecewise rigid motion. Given the sequence, the generative model which best describes an articulated object and a background is learnt in an unsupervised manner. The generative model for a layered representation describes the scene as a composition of layers.

Files

• Master thesis CTU–CMP–2007–03.pdf
• Presentation