Spatio-temporal constraint for fast face tracking in movies

In this paper, a new unified framework of the face tracking is presented. It is based on a new tracking-by-detection tracker. Three different face detectors are utilized in the tracker. By exploiting the spatio-temporal constraint between the intra and inter frames, a Bayesian formulation is proposed to merge different detections from the three detectors and link the faces into tracks. In addition, a salient region in a frame is found by employing context prior knowledge. The tracking procedure is efficiently accelerated because that the three sliding-window based detectors scan in the smaller salient region instead of a whole frame. Our method is evaluated on the standard Hannah dataset, which contains a feature-length movie. The performance is demonstrated to match or exceed the state-of-the-art. Furthermore, our system is much faster than previous methods.