Skip to Main Content
The goal of comparing two three-dimensional (3D) facial models is to decide if they belong to a certain subject or not. Selecting discriminative feature is challenging when taking expressions into consideration, since intra-personal and inter-personal disparities exist simultaneously in the residue of two finely registered models. The intra-personal deformation caused by expressions is the main factor decreasing the recognition performance of matching algorithms. A proposed multi-scale model addresses this issue by approximating expression changes in a two-step process. Employing manifold harmonics, the model decomposes a face into three components according to its frequency domains. Low-frequency components are used to estimate the intra-personal residue in the second step. Experiments conducted on the face recognition grand challenge version 2 dataset demonstrate the benefit of our scheme over the rigid matching method. An extensive performance evaluation shows that the proposed approach is promising for dealing with expression distortions in 3D face recognition.