Parametric model-based motion segmentation using surface selection criterion

This paper presents a new framework for the motion segmentation task, which includes an algorithm capable of addressing the important issue of the inter-relationships between data segmentation, model selection, and noise scale estimation. In this algorithm, we have incorporated our newly proposed model selection criterion named Surface Selection Criterion. The presented algorithm simultaneously selects the correct motion model, while finding the scale of the noise and performing the segmentation task. As a result, the estimated motion parameters and the final segmentation results are accurate. The algorithm is tested for motion segmentation of synthetic and real video data containing multiple objects undergoing different types of motion. Our results also show that the proposed algorithm is capable of detecting occlusion and degeneracy.