Unsupervised co-segmentation for 3D shapes using iterative multi-label optimization

This paper presents an unsupervised algorithm for co-segmentation of a set of 3D shapes of the same family. Taking the oversegmentation results as input, our approach clusters the primitive patches to generate initial guess. Then, it iteratively builds a statistical model to describe each cluster of parts from previous estimation, and employs the multi-label optimization to improve the co-segmentation results. In contrast to the existing “one-shot” algorithms, our method is superior in that it can improve the cosegmentation results automatically. The experimental results on Princeton segmentation benchmark demonstrate that our approach is able to co-segment the 3D shapes with significant variability and achieves comparable performance to the existing supervised algorithms and better performance than the unsupervised ones.