Spatial alignment network for facial landmark localization

Huifang Li, Yidong Li, Junliang Xing, Hairong Dong

Abstract: Facial Landmark Localization (FLL) on unconstrained images still remains challenging as they poses complex variation in face spatial structure and appearance. To address this problem, we propose a Spatial Alignment Network (SAN), which consist of two modules, like the transformation sub-network and the estimation sub-network. In the first module, we propose two methods to achieving spatial transformation, one is the handcrafted method which can ensure model stability and the other is the learning-based method which is efficient and flexible. In the second module, we add an attention layer in the deep CNN to enhance the importance of discriminative features and obtain more accurate results. Through extensive experiments, our model achieves good performance on several public challenging datasets.