Exploring manifold learning algorithms for shape-based 3D model retrieval

A distance measure, along with a shape feature, is the most critical components in a shape-based 3D model retrieval system. While various distance measures have been proposed, a single distance measure can’t handle variations per database. We have previously reported that a non-linear manifold learning algorithm that trained unsupervised by a set of 3D models database could significantly improve retrieval performance for a pair of shape features [Ohbuchi06]. In this paper, we experimentally explore the approach further. We do so by applying six different kinds of linear and non-linear manifold learning algorithm on five shape features and their multi-resolution versions. The experiments showed that a set of non-linear, local manifold learning algorithms works quite well when combined with five different kind of shape features and their multiresolution versions.