Unconstrained Face Detection

Face detection, as the first step in automatic facial analysis, has been well studied over the past two decades. However, challenges still remain for face detection in unconstrained scenarios, such as arbitrary pose variations and occlusions. In this paper, we propose a method to address these challenges in unconstrained face detection. First, a new type of image feature, called Normalized Pixel Difference (NPD) is proposed. NPD feature is computed as the difference to sum ratio between any two pixel intensity values, inspired by the Weber Fraction in experimental psychology. Besides its computational efficiency, the NPD feature has several desirable properties, such as scale invariance, boundedness, and ability to reconstruct the original image. Second, we develop a method for learning the optimal subset of NPD features and their combinations via regression trees, so that complex face manifolds can be partitioned by the learned rules. This way, only a single cascade classifier is needed to handle unconstrained face detection. The proposed face detector is robust in handling pose, occlusion, illumination, blur and low image resolution. Experimental results on three public face datasets (FDDB, GENKI, and CMU-MIT) show that the proposed method outperforms the state-of-the-art methods reported to date in detecting unconstrained faces with arbitrary pose variations and occlusions in cluttered scenes.