Efficient frame-sequential label propagation for video object segmentation

Abstract

In this work, we present an approach for segmenting objects in videos taken in complex scenes. It propagates initial object label through the entire video by a frame-sequential manner where the initial label is usually given by the user. The proposed method has several contributions which make the propagation much more robust and accurate than other methods. First, a novel supervised motion estimation algorithm is employed between each pair of neighboring frames, by which a predicted shape model can be warped in order to segment the similar color around object boundary. Second, unlike previous methods with fixed modeling range, we design a novel range-adaptive appearance model to handle the tough problem of occlusion. Last, the paper gives a reasonable framework based on GraphCut algorithm for obtaining the final label of the object by combining the clues from both appearance and motion. In the experiments, the proposed approach is evaluated qualitatively and quantitatively with some recent methods to show it achieves state-of-art results on multiple videos from benchmark data sets.