Pose-Invariant Probabilistic Face Recognition

Modern deep learning-based face recognition systems perform at or above human levels on high quality frontal imagery, but their performance suffers significantly on more operationally relevant imagery with confounding factors such as extreme pose.  This performance degradation is based on two primary shortcomings: poor match between training and test-time image data, and failure to model inherent uncertainty present in extreme pose and otherwise low-quality image data. The proposed face recognition framework achieves high performance on challenging imagery through the novel combination of two primary technical components: 3D facial data augmentation and deep network-based match confidence estimation.  First, the recognition system is optimized for operational data by using 3D data augmentation to bridge the gap between the pose distributions of the training and test sets, made possible through the use of VSI's proven 3D face shape estimation and rendering framework. Second, the proposed recognition system is designed to improve performance at low false alarm rates through the use of probabilistic matching and fusion, made possible by representing the set of possible identities present in a given image as a multivariate probability distribution.