A Novel 3D Facial Expression Recognition Approach Based on Tensor Distance
Qian Li, Qiuqi Ruan · 2018
I65 this paper, a novel tensor subspace learning algorithm for 3D facial expression recognition-Tensor Distance based Orthogonal Tensor Marginal Fisher Analysis (TD-OTMFA) is proposed. In addition, we also improve it and get its weighted version-WTD-OTMFA. Firstly, five different kinds of maps including Depth Map (DM), three Normal Maps, and Shape Index Map (SIM) are adopted to describe 3D facial expressions. Then TD-OTMFA and WTD-OTMFA are used to learn the low dimensional representation of the expressions. Compared with the traditional subspace learning algorithms, TD-OTMFA directly processes samples into tensors so that the spatial structure can be preserved and avoids the curse of dimensionality. Moreover, the introduction of Tensor Distance (TD) makes it no longer constrained by the orthogonality assumption as in Euclidean space. That's, it considers the correlation among different coordinates when measuring distances among samples. WTD-OTMFA takes full account of the local structure of faces and emphasizes the contributions of different regions of faces. This process makes it more discriminative. Experiments conducted on the BU-3DFE dataset demonstrate that our proposed method outperforms the-state-of-the-art ones.