3D shape representation with spatial probabilistic distribution of intrinsic shape keypoints
This study developed a new shape-based 3D model descriptor based on the D2 shape descriptor developed by Osada, et al of Princeton University. Shape descriptors can be used to measure dissimilarity between two 3D models. In this work, we advance it by proposing a novel descriptor D2a. In our method, N pairs of faces are randomly chosen from a 3D model, with probability proportional to the area of the face. The ratio of the smaller area over the larger area is computed and its frequency stored, generating a frequency distribution of N ratios which is stored as the second dimension of a 2D array, while the first dimension contains the frequency distribution of distances of randomly generated point pairs (the D2 distribution). The resulting descriptor, D2a, is a two-dimensional histogram that incorporates two shape features: the ratio of face areas and the distance between two random points.