Non-rigid 3D Object Retrieval with a Learned Shape Descriptor

Non-rigid 3D objects are difficult to distinguish due to the structural transformation and noises. In this paper we develop a novel method to learn a discriminative shape descriptor for non-rigid 3D object retrieval. Compact low-level shape descriptors are designed from spectral descriptor, and the non-linear mapping of low level shape descriptors is carried out by a Siamese network. The Siamese network is trained to maximize the inter-class margin and minimize the intra-class distance. With an appropriate network hierarchy, we extract the last layer of the successfully trained network as the high-level shape descriptor. Furthermore, we successfully combine two low-level shape descriptors, based on the Heat Kernel Signature and the Wave Kernel Signature, and test the method on the benchmark dataset SHREC'14 Non-Rigid 3D Human Models. Experimental results show our method outperforms most of the existing algorithms for 3D shape retrieval.