Unsupervised learning from a corpus for shape-based 3D model retrieval

Abstract

Arguably the most important issues in shape-based 3D model retrieval are methods to extract powerful yet compact shape features and methods to properly and promptly compare the shape features. In this paper, we explore a method to improve feature distance computation by employing unsupervised learning of the subspace of 3D shape features from a corpus. We employ an algorithm called Laplacian Eigenmaps proposed by Belkin, et al. to learn a manifold spanned by shape features of 3D models in the corpus. The learned manifold is approximated by an RBF network, onto which shape features are projected. Distances among shape features can then be computed effectively on the learned manifold. We combine this learning-based distance-computation method with a method to extract multiresolution shape features proposed by Ohbuchi, et al. Our experimental evaluation showed that the proposed method could significantly improve retrieval performance. Learning alone improved performance of two shape features we tried by about 5%. A combination of learning and multiresolution shape feature allowed about 10% gain in performance. As an example, the trained, multiresolution version of the SPRH gained 10% over the original single resolution, untrained SPRH shape feature.