Face Recognition and Pose Estimation with Parametric Linear Subspaces [chapter]

Kazunori Okada, Christoph von der Malsburg
2008 Studies in Computational Intelligence  
We present a general statistical framework for modeling and processing head pose information in 2D grayscale images: analyzing, synthesizing, and identifying facial images with arbitrary 3D head poses. As a basic component, LPCMAP model offers a compact view-based datadriven model with bidirectional mappings between face views and 3D head angles. We call a mapping from face to pose analysis mapping and that from pose to face synthesis mapping. A model matching is also defined by an
more » ... hesis chain that concatenates the two mappings. Such a mapping-based model implicitly captures 3D geometric nature of the problem without explicitly reconstructing 3D facial structure from data. The model is learned by using efficient PCA and SVD algorithms resulting in linear functional forms. They are however only locally valid due to the linear design. We further extend this local model to mitigate the shortcomings. PPLS model extends the LPCMAP for covering a wider pose range by piecing together a set of LPCMAPs. Multiple-PPLS model further extends the PPLS for generalizing over different individuals. These proposed models are applied to solve pose estimation and animation by using the analysis and synthesis mappings, respectively. A novel pose-insensitive face recognition framework is also proposed by exploiting the PPLS model to represent each known person. In our recognition framework, the model * The corresponding author can also be reached at 3641 Lavell Drive Los Angeles, CA 90065, kazokada@earthlink.net. Main part of the presented work was conducted while both authors were at University of Southern California, Los Angeles USA. matching with the PPLS models provides a flexible pose alignment of model views and input faces with arbitrary head poses, making the recognition invariant against pose variation. Implementations of the proposed models are empirically evaluated with a database of various views of 20 people rendered from Cyberware-scanned 3D face models. The results demonstrated sub-degree pose estimation and sub-pixel shape synthesis accuracy, as well as high degree of generalization to unseen head poses within ±50 degree range of full 3D head rotation. For recognition and interpersonalized pose estimation, the results also indicated robustness against unseen head poses and individuals while compressing the data by a factor of 20 and more.
doi:10.1007/978-3-540-76831-9_3 fatcat:lkn3mgastbb4bgwztkboc5rowe