Probabilistic Motion Segmentation of Videos for Temporal Super Resolution

A novel scheme is proposed for achieving motion segmentation in low-frame rate videos, with application to temporal super resolution. Probabilistic generative models are commonly used to perform unsupervised motion segmentation in videos. While they provide a general and elegant framework, they are hampered by severe local minima problems and often converge to inaccurate solutions, when there are more than one foreground object in videos. This paper proposes a scheme, where discriminative global constraints are enforced in combination with generative learning, to overcome the local minima problems. We demonstrate the effectiveness of the proposed scheme by learning the appearances and motions of multiple objects from a low frame rate video with a small number of frames.