A Novel HKS Based Feature Extraction Algorithm

Heat Kernel Signature (HKS) is a powerful tool for shape analysis for its multi useful properties and has been successfully used in many correspondence tasks. However, the shape's feature detection is an empirical way since HKS depends on the time scale, which is not the intrinsic property of the shape. In order to eliminate the effects of time ambiguity, a novel HKS based feature extraction algorithm is introduced in this paper. It extracts points that have the largest changes in HKS value and take them as feature points since the energy of heat kernels changes according to Gaussian curvature, which is an intrinsic value. The extracted feature points can well reflect the shape's sharp features and thus can be used in many correspondence related tasks. Experiments are performed on isometric shapes and partial shapes in the TOSCA dataset. The results show that the proposed feature extraction method achieves good robustness and effectiveness for these tasks.