Probabilistic Fingerprints for Shapes

Abstract

We propose a new probabilistic framework for the efficient estimation of similarity between 3D shapes. Our framework is based on local shape signatures and is designed to allow for quick pruning of dissimilar shapes, while guaranteeing not to miss any shape with significant similarities to the query model in shape database retrieval applications. Since directly evaluating 3D similarity for large collections of signatures on shapes is expensive and impractical, we propose a suitable but compact approximation based on probabilistic fingerprints which are computed from the shape signatures using Rabin s hashing scheme and a small set of random permutations. We provide a probabilistic analysis that shows that while the preprocessing time depends on the complexity of the model, the fingerprint size and hence the query time depends only on the desired confidence in our estimated similarity. Our method is robust to noise, invariant to rigid transforms, handles articulated deformations, and effectively detects partial matches. In addition, it provides important hints about correspondences across shapes which can then significantly benefit other algorithms that explicitly align the models. We demonstrate the utility of our method on a wide variety of geometry processing applications.