Explicit Occlusion Detection based Deformable Fitting for Facial Landmark Localization

This paper addresses the problem of facial landmark localization on partially occluded faces. We proposes an explicit occlusion detection based deformable fitting model for occluded landmark localization. Most recent shape registration methods apply landmark local search and attempt to simultaneously minimize both the model error and localization error. However, if part of the shape is occluded, those methods may lead to misalignment. In this paper, we introduce regression based occlusion detection to restrict the occluded landmarks' error propagation from passing to the overall optimization. Assuming the parameter model being Gaussian, we propose a weighted deformable fitting algorithm that iteratively approaches the optima. Experimental results in our synthesized facial occlusion database demonstrate the advantage of our method.