A new frame for exemplar-based shape clustering

Unsupervised clustering of objects is often needed for image and video summarization, tracking and segmentation. Shape, as fundamental representation of objects, is hard to do clustering task since usual clustering algorithms need quantitative features which are very hard to extract in shapes. In this paper, we proposed a novel approach to shape clustering. To overcome the difficulty of extracting feature vectors in the unsupervised task of shape clustering, we provide a novel method to iteratively learn the best cluster centers. We modify the frame of fuzzy clustering algorithm by effectively choosing representative exemplars. Unsupervised categorization by identifying a subset of representative exemplars can be efficiently performed by our new framework. When applied to some famous shape datasets, our method achieves a much lower reconstruction error.