The performance of still-to-video face recognition (FR) systems can declinesignificantly because faces captured in the unconstrained operational domain(OD) have a different underlying data distribution compared to faces capturedunder controlled conditions in the enrollment domain (ED). This is particularlytrue when individuals are enrolled to the system using a single referencestill. To improve the robustness of these systems, it is possible to augmentthe gallery set by generating synthetic faces based on the original still.However, without the OD knowledge, many synthetic faces must be generated toaccount for all possible capture conditions. FR systems may therefore requirecomplex implementations and yield lower accuracy when training on less relevantimages. This paper introduces an algorithm for domain-specific face synthesis(DSFS) that exploits the representative intra-class variation informationavailable from the OD. Prior to operation (during camera calibration), acompact set of faces from unknown persons appearing in the OD is selectedthrough clustering in the captured condition space. The domain-specificvariations of these faces are projected onto the reference still of eachindividual by integrating an image-based face relighting technique inside a 3Dreconstruction framework. A compact set of synthetic faces is generated underthe OD capture conditions. In a particular implementation based on sparserepresentation classification, the synthetic faces generated with the DSFS areemployed to form a cross-domain dictionary where the dictionary blocks combinethe original and synthetic faces of each individual. Experimental resultsobtained with the Chokepoint and COX-S2V datasets reveal that augmenting thegallery set using the DSFS approach provide a higher level of accuracy comparedto state-of-the-art methods, with only a moderate increase in its complexity.
展开▼
