Retrieval of non-rigid 3D shapes from multiple aspects

As non-rigid 3D shape plays increasingly important roles in practical applications, this paper addresses its retrieval problem by considering three aspects: shape representation, retrieval optimization, and shape filtering. (1) For shape representation, two kinds of features are considered. We first propose a new integration kernel based local descriptor, and then an efficient voting scheme is designed for shape representation. Besides, we also study the commute times as shape distributions, which grasp the spatial shape information globally. Both of them capture shape information from different viewpoints based on the same embedding basis. (2) We then study the typical problem of retrieval optimization. Prior works show poor stability under different similarity windows. To deal with this deficiency, we propose to model the problem as a distance mapping on a graph in spectral manifold space. (3) Usually, for each retrieval input, a list is returned and there may be lots of irrelevant results. We develop an algorithm to filter them out by combining multiple kernels. Finally, three public datasets are employed for performance evaluation and the results show that the studied techniques have contributed a lot in promoting the recognition rate of non-rigid 3D shapes.