Face detection by structural models

Despite the successes in the last two decades, the state-of-the-art face detectors still have problems in dealing with images in the wild due to large appearance variations. Instead of leaving appearance variations directly to statistical learning algorithms, we propose a hierarchical part based structural model to explicitly capture them. The model enables part subtype option to handle local appearance variations such as closed and open month, and part deformation to capture the global appearance variations such as pose and expression. In detection, candidate window is fitted to the structural model to infer the part location and part subtype, and detection score is then computed based on the fitted configuration. In this way, the influence of appearance variation is reduced. Besides the face model, we exploit the co-occurrence between face and body, which helps to handle large variations, such as heavy occlusions, to further boost the face detection performance. We present a phrase based representation for body detection, and propose a structural context model to jointly encode the outputs of face detector and body detector. Benefit from the rich structural face and body information, as well as the discriminative structural learning algorithm, our method achieves state-of-the-art performance on FDDB, AFW and a self-annotated dataset, under wide comparisons with commercial and academic methods.