Accurate and Efficient Similarity Search on 3D Objects Using Point Sampling, Redundancy, and Proportionality

With fast evolving resources for 3D objects such as the Protein Data Bank (PDB) or the World Wide Web, new techniques, so-called similarity models to efficiently and effectively search for these 3D objects become indispensible. Invariances w.r.t. specific geometric transformations such as scaling, translation, and rotation are important features of similarity models. In this paper, we focus on rotation invariance. We first propose a new method of representing objects more accurately in the context of rotation invariance than the well-known voxelization technique.In addition, we extend existing feature-based similarity models by proposing a new spherical partitioning of the data objects based on proportionality and redundancy, and generalizing an existing method for feature extraction. A broad experimental evaluation compares our method with existing methods in terms of accuracy and efficiency. In particular, we experimentally confirm that our point sampling method is better suited to represent 3D objects in the context of rotation invariance than voxelized representations. In addition, we empirically show that our new similarity model significantly outperfoms competitive rotation invariant models in terms of accuracy as well as efficiency.