From co-saliency detection to object co-segmentation: A unified multi-stage low-rank matrix recovery approach

Abstract

Object co-segmentation aims to identify and segment the common objects among a set of similar images. Although various explorations have been done for the topic, two major problems still remain: (1) How to mitigate the influence of background disturbance of each image when we detect the common objects. (2) How to leverage common information of the image set optimally. To overcome the two problems, we resort to co-saliency detection and propose a novel framework, which utilizes multi-stage low-rank matrix recovery to eliminate the background and identify the common foregrounds. To address the first problem, we firstly use a conventional saliency detection model to get saliency maps of each image as initialization rather than directly dealing with all the images together; to address the second problem, we adopt low-rank matrix recovery to constrain the common foregrounds as the low-rank part, while the background interferences corresponds to the sparse noises. Besides, an effective refinement method is proposed to recover the spatial relationships among the segments. The extensive experiments show the proposed model can effectively leverage the homogeneous information among the image class and provide promising co-segmentation performance.