COMPACT PART-BASED SHAPE SPACES FOR DENSE CORRESPONDENCES OLIVER BURGHARD, UNIVERSITY

Abstract. We consider the problem of establishing dense correspondences within a set of related shapes of strongly varying geometry. For such input, traditional shape matching approaches often produce unsatisfactory results. We propose an ensemble optimization method that improves given coarse cor-respondences to obtain dense correspondences. Following ideas from minimum description length approaches, it maximizes the compactness of the induced shape space to obtain high-quality correspondences. We make a number of improvements that are important for computer graphics applications: Our ap-proach handles meshes of general topology and handles partial matching be-tween input of varying topology. To this end we introduce a novel part-based generative statistical shape model. We develop a novel analysis algorithm that learns such models from training shapes of varying topology. We also provide a novel synthesis method that can generate new instances with varying part lay-outs and subject to generic variational constraints. In practical experiments, we obtain a substantial improvement in correspondence quality over state-of-the-art methods. As example application, we demonstrate a system that learns shape families as assemblies of deformable parts and permits real-time editing with continuous and discrete variability. (a) shape collection (b) coarse segmentation (c) dense correspondences 1