Non-rigid 3D Shape Retrieval via Sparse Representation

Shape descriptor design is an important but challenging problem for non-rigid 3D shape retrieval. Recently, bagof-words based methods are widely used to integrate a model’s local shape descriptors into a global histogram. In this paper, we present a new method to pool the local shape descriptors into a global shape descriptor by means of sparse representation. Firstly, we employ heat kernel signature (HKS) to depict the multi-scale local shape. Then, for each model in the training dataset, we take the HKSs corresponding to its mesh vertices to serve as training signals, and thus an over-complete dictionary can be learned from them. Finally, the HKSs of each 3D model are sparsely coded based on the learned dictionary, and such sparse representations can be further integrated to form an object-level shape descriptor. Moreover, we conduct extensive experiments on the state-of-the-art benchmarks, wherein comprehensive evaluations state our method can achieve better performance than other bag-of-words based approaches.