UC Riverside

Existing face recognition algorithms perform well under controlled conditions; however, in real-world scenarios, where variations in pose, illumination, scale, rotation, and expression are encountered, they are unable of performing accurately and/or efficiently. As an example, in the case of the Boston marathon bombings on April 15th, 2013, although the suspects' face images existed in government databases (e.g., DMV), the face recognition algorithms were unable to identify the suspects before they were identified by other means. Motivated by the demand of reliable face recognition techniques for real-world scenarios, we are tasked with identifying an unknown face, or verifying a match between a pair of unknown faces. The process of doing so requires overcoming many uncertainties in real-world images including pose and scale variation, misalignment, varying illumination conditions, etc. Towards this goal, two new recognition frameworks, a graph-based probabilistic pose estimation classifier, and a hash-based retrieval algorithm were designed to accurately perform face recognition in large-scale real-world applications. In addition, a novel real-time video-based classification system was developed for robust vehicle classification.