Adaptive Non-local Affinity Graph for Unsupervised Image Segmentation

In this paper, we present an adaptive superpixel-based graph construction method for unsupervised image segmentation. The empirical findings in this study provide a new understanding of applying non-local image patches to image segmentation. We demonstrate that high-level segmentation cues can be revealed by non-local similar patches, and based on it the similarity between superpixels overlapped with these similar patches can be adaptively constructed. The experimental result on the Berkeley Segmentation Database demonstrates the effectiveness and superiority of the proposed algorithm.