Face alignment by Component Adaptive Mechanism

Abstract The new cascaded shape regression architecture proposed in this paper is actually an algorithm by Component Adaptive Mechanism (CAM) to cope with unconstrained face alignment. CAM divides the process of face alignment into two parts: the updating process of face box and the cascaded shape regression process by Component Adaptive Mechanism. The former process first adjusts face box by training different classifiers to estimate its transformation parameters and thereby outputs more accurate initialized shape. The latter process uses Component Adaptive Mechanism to fuse results of different domain-specific regressors to further update the shape. The major innovation of CAM is characterized by its fault-tolerated mechanism which is shown in the following two aspects. (1) A probability-based fern classifier is adopted in the partition of the optimization space into multiple domains of homogeneous descent, which not only endows the algorithm with the fault-tolerance mechanism but also augments the available training set of each domain. (2) A training strategy based on dominant set approach is used to train a stronger domain-specific regressor by dynamically adjusting the weight of objective function corresponding to different shapes and therefore regressors derived from the training are equipped with fault-tolerance ability. Conducted on such public image datasets as AFLW-full (19-pts), COFW (29-pts) and 300-W (68-pts), experiments show that the proposed CAM: (1) can deal with the problem of face detection and face alignment simultaneously; (2) is superior to existing algorithms in solving face alignment problems with extreme variations in pose, expression, illumination and partial occlusion.