Unsupervised video co-segmentation based on superpixel co-saliency and region merging

Abstract

Nowadays, fully unsupervised video object segmentation is still a challenge in computer vision. Furthermore, it is more difficult to segment the object from a set of clips. In this paper, we propose an unsupervised and on-line method that efficiently segments common objects from a set of video clips. Our approach is based on the hypothesis, that common or similar objects in multiple video clips are salient, and they share similar features. At first, we try to find out the regions in every clip which are salient and share similar features by proposing a new co-saliency scheme based on superpixels. Then, the most salient superpixels are chosen as the initial object marker superpixels. Starting from these superpixels, we merge neighboring and similar regions, and segment out the final object parts. The experimental results demonstrate that the proposed method can efficiently segment the common objects from a group of video clips with generally lower error rate than some state-of-the-art video co-segmentation methods.