Output details

TENSOR-BASED FEATURE REPRESENTATION WITH APPLICATION TO MULTIMODAL FACE RECOGNITION

Originality: Facial expression capture has become increasingly important for facial animation in film production (e.g., Avatar, The Adventures of Tintin). To capture a facial expression, data must be collected from multiple difference sources. Algorithms are urgently needed for the integration of such data for applications including facial recognition, feature extraction and retargeting (reusing) facial animation on a new animation character (avatar). This paper addresses this fundamental problem at a theoretical level and presents a novel geometric descriptor which lays the basis for integrating data from different formats under the same framework. A new N-Dimensional Principle Component Analysis (ND-PCA) classifier is also presented for facial animation and recognition.

Significance: This research sets a theoretical foundation for multi-mode facial feature representation. This new feature descriptor integrates all important information from human faces into a single compact form which will significantly improve the efficiency and accuracy of facial animation and facial expression based human computer interaction. This paper also presents a common baseline classifier (ND-PCA). The performance of all other methods can be numerically evaluated by comparing against this baseline classifier, no matter what data structure they use.

Rigour: Various numerical properties of the proposed feature descriptor have been proven both theoretically and experimentally. To evaluate the effectiveness of the proposed ND-PCA classifier, this work also gives a theoretical estimation of the upper error bound, and proved that it is linear optimal. The experimental results indicated that the proposed feature descriptor could effectively improve the computation accuracy.