Lightweight Binary Voxel Shape Features for 3D Data Matching and Retrieval

摘要

This paper proposes several lightweight local 3D shape features for 3D voxel data that yield compact binary feature vectors. These features are inspired by compact binary features for 2D image, namely, Local Binary Pattern (LBP) [22], BRIEF [6] and ORB [26]. In addition to being compact, extraction of proposed 3D features is inexpensive. Furthermore, these binary feature vectors are very efficient to compare, as their distance in Hamming space can be computed very efficiently. Our experimental evaluation of these features in a shape-based 3D model retrieval setting showed that some of these 3D binary features perform competitively to some of existing features. Depending on benchmark database, proposed features are somewhat less accurate than or about as accurate as the state-of-the-art 3D shape features. However, memory footprint is much more compact, at about 1/10 of the non-binary 3D shape features having comparable retrieval accuracy.