Video object segmentation based on superpixel trajectories

Abstract

In this paper, a video object segmentation method utilizing the motion of superpixel centroids is proposed. Our method achieves the same advantages of methods based on clustering point trajectories, furthermore obtaining dense clustering labels from sparse ones becomes very easy. Simply for each superpixel the label of its centroid is propagated to all its entire pixels. In addition to the motion of superpixel centroids, histogram of oriented optical flow, HOOF, extracted from superpixels is used as a second feature. After segmenting each object, we distinguish between foreground objects and the background utilizing the obtained clustering results.