Delving Deep Into Coarse-to-Fine Framework for Facial Landmark Localization

Abstract

In this paper we proposed a 4-stage coarse-to-fine framework to tackle the facial landmark localization problem in-the-wild. In our system, we first predict the landmark key points on a coarse level of granularity, which sets a good initialization for the whole framework. Then we group the key points into several components and refine each component with local patches cropped within them. After that we further refine them separately. Each key point is further refined with multi-scale local patches cropped according to its nearest 3-, 5-, and 7-neighbors respectively. The results are fused by an attention gate network. Since a different key-point configuration is adopted in our labeled dataset, a linear transformation is finally learned with the least square approximation to adapt our predictions to the competition's task.