Noise-aware co-segmentation with local and global priors

Image segmentation is a long-standing challenge in image and video processing. The method of co-segmentation aims at discovering common foreground object shared in image set. The traditional co-segmentation methods usually assume that all images should contain the target object. In this paper, we perform co-segmentation by first refining the image set. To this end, we propose to use attentiveness score, which is built upon the semantic proposals to identify the target object. We further filter out the noisy images using affinity propagation clustering. Then, both local and global shape priors are computed from the cleaned image set. The local prior can accurately estimate the foreground boundary, and the global prior supervises the pose and viewpoint of target object. These priors are optimized via dense correspondence mapping. Finally, we perform co-segmentation by minimizing an energy function. Experiments on three testbeds including Graz02, Internet images and MSRC object dataset, demonstrate that the proposed method outperforms the state-of-the-art co-segmentation methods.