As the usage of 3D models increases, so does the importance of developing accurate 3D shape retrieval algorithms. Most prominently, shape descriptors are used to describe the geometric and topological properties of objects and compared to determine two objects similarity. They are split into two categories global and local. As local descriptors are in general more invariant to rotation, translation, and scaling, and can additionally be applied to articulated models and partial matching problems, many have been proposed. However, these descriptors are often evaluated independently and on different datasets, making them difficult to compare. Using the SHREC 2011 Shape Retrieval Contest of Non-rigid 3D Watertight Meshes dataset, we systematically evaluate a collection of local shape descriptors. We choose to apply them to the bag-of-words paradigm, where each object is represented as a histogram counting occurrences of each word of a visual dictionary. In addition, the role of vocabulary size and number of sample points taken from each object in performance is assessed. Salient point detection methods are applied to determine if the number of sample points can be decreased without sacrificing accuracy. Finally, information from two local descriptors is combined in a number of ways and changes in performance are investigated.
Citation: NIST Interagency/Internal Report (NISTIR) - 7812
NIST Pub Series: NIST Interagency/Internal Report (NISTIR)
Pub Type: NIST Pubs
3D shape retrieval, local features, bag-of-words, evaluation