Temporally consistent region based video segmentation

Abstract

This thesis addresses the problem of segmenting a video sequence in a temporally consistent fashion, so that the labels assigned to particular region remain the same throughout the sequence. Energy minimisation, and in particular �-expansion based label propagation, is a popular approach for addressing video segmentation. In this approach, labels of most similar object classes are assigned to pixels such that the energy of final libelling is minimised. Two major limitations of this approach are not being able to handle non class based segmentation and appearance of new objects in the current frame. In this thesis, we developed a novel method for label propagation in the forward direction using the �-expansion method.