Unconstrained Face Recognition

Thumbnail Image


umi-umd-1782.pdf (5.9 MB)
No. of downloads: 3217

Publication or External Link






Although face recognition has been actively studied over the past decade, the state-of-the-art recognition systems yield satisfactory performance only under controlled scenarios and recognition accuracy degrades significantly when confronted with unconstrained situations due to variations such as illumintion, pose, etc. In this dissertation, we propose novel approaches that are able to recognize human faces under unconstrained situations.

Part I presents algorithms for face recognition under illumination/pose variations. For face recognition across illuminations, we present a generalized photometric stereo approach by modeling all face appearances belonging to all humans under all lighting conditions. Using a linear generalization, we achieve a factorization of the observation matrix consisting of face appearances of different individuals, each under a different illumination. We resolve ambiguities in factorization using surface integrability and symmetry constraints. In addition, an illumination-invariant identity descriptor is provided to perform face recognition across illuminations. We further extend the generalized photometric stereo approach to an illuminating light field approach, which is able to recognize faces under pose and illumination variations.

Face appearance lies in a high-dimensional nonlinear manifold. In Part II, we introduce machine learning approaches based on reproducing kernel Hilbert space (RKHS) to capture higher-order statistical characteristics of the nonlinear appearance manifold. In particular, we analyze principal components of the RKHS in a probabilistic manner and compute distances such as the Chernoff distance, the Kullback-Leibler divergence between two Gaussian densities in RKHS.

Part III is on face tracking and recognition from video. We first present an enhanced tracking algorithm that models online appearance changes in a video sequence using a mixture model and produces good tracking results in various challenging scenarios. For video-based face recognition, while conventional approaches treat tracking and recognition separately, we present a simultaneous tracking-and-recognition approach. This simultaneous approach solved using the sequential importance sampling algorithm improves accuracy in both tracking and recognition. Finally, we propose a unifying framework called probabilistic identity characterization able to perform face recognition under registration/illumination/pose variation and from a still image, a group of still images, or a video sequence.