Propagating multi-class pixel labels throughout video frames

The effective propagation of pixel labels through the spatial and temporal domains is vital to many computer vision and multimedia problems, yet little attention have been paid to the temporal/video domain propagation in the past. Previous video label propagation algorithms largely avoided the use of dense optical flow estimation due to their computational costs and inaccuracies, and relied heavily on complex (and slower) appearance models. We show in this paper the limitations of pure motion and appearance based propagation methods alone, especially the fact that their performances vary on different type of videos. We propose a probabilistic framework that estimates the reliability of the sources and automatically adjusts the weights between them. Our experiments show that the “dragging effect” of pure optical-flow-based methods are effectively avoided, while the problems of pure appearance-based methods such the large intra-class variance is also effectively handled.