Supervoxel Based Video Segmentation

This paper describes the development of video segmentation and its related field. Atthe conclusion on the basis of previous results and combined with the understanding of theproblem, we propose the supervoxel segmentation using partially absorbing random walksalgorithm. Experimental results demonstrate that our method achieves better performancecompared to the state-of-the-art supervoxel algorithms.The main contribution of this paperas follows:1) This paper presents a supervoxel generation algorithm based on partially absorbingrandom walks to get supervoxels. The proposed spatial-temporal framework performs thesupervoxel segmentation in two adjacent frames, rather than processing the entire video clip.It greatly simplifies the computational complexity, memory and other resource consumptionis also significantly reduced. With the propagation of each frame segmentation results, wegenerate supervoxels with more smoothness in both spatial and temporal area.(2) We have established a novel Laplacian optimize the structure between twoconsecutive frames. This Laplacian matrix fully takes the advantage of appearance featuresand motion cues, which is a key factor for obtaining consistent and accurate supervoxels.(3) The partially absorbing random walk algorithm overcomes the misleading problemof traditional random walk, and it can effectively consider the global information whichgenerates nice performance for weak boundaries and complex textures area, overcomingthe problem of supervoxels with overlapping the boundaries of objects. Therefore, ouralgorithm can also be sufficient to maintain the consistency even in the area of weakboundaries and complex textures.