Visibility-aware part model for robust facial point detection

Abstract

Abstract In this paper, we present a visibility-aware part model for precise facial point detection under occlusions, based on the pictorial structure model. The pictorial structure is a computationally efficient framework for modelling part-based objects. We introduce a binary part visibility term for the model to describe the occlusion state of each part, which can determine which facial points are occluded. The introduction of the term enhances the representation power of the model especially for the occlusions. The combination of the structure constrains and the powerful appearance model makes the model more robust and reduces the possibility of model crashing some extent. Experimental results show that our proposed model can detect facial feature points accurately and robustly under occlusions.