Site-adaptive face recognition

While the state-of-the-art face recognition algorithms are designed with the goal of reliably recognizing faces under uncontrolled imaging conditions, the performance of these face recognizers varies in the real-world applications, depending on how much the face appearance statistics in the testing data matches with those in the training data. Assuming the imaging condition is not subject to frequent changes at a particular application site where the face recognition systems are deployed, we propose to do site adaptation for a generic face recognizer based on a small adaptation data set captured at the site. Based on an OSFV face recognizer with Gabor features selected by Adaboost algorithm, we propose several site adaptation methods at the feature level and at the model level. Our experiment results show that the proposed site adaptation approaches can boost the performance of our generic face recognition algorithm based on a small adaptation dataset acquired from the site with a different imaging condition.