Perceptually friendly shape decomposition by resolving segmentation points with minimum cost

Organizing objects in terms of parts constructs a natural shape descriptor and provides a foundation for many shape-related applications, such as object detection and tracking. In this paper, we propose a part decomposition scheme that focuses on analyzing the relations between part cuts and segmentation points. Aiming at obtaining perceptually friendly parts, we take the human perception behavior into account when extracting candidate part cuts and defining their costs. The decomposition is to find a subset from candidate part cuts with minimum cost, while guarantee that all segmentation points are resolved. The theoretical and experimental results demonstrate that the proposed method is robust to occlusion, part movement and boundary distortion. We also show that our method can predict most of highly salient part cuts, and have a good performance in hand gesture recognition.

► The optimization problem simulates interplay among various visual factors.
► Based on DCE, our method satisfies the requirements of object-based decomposition.
► Relation matrices defined based on relevance measure are friendly to human vision.