Stacked hourglass (HG) networks have been successfully applied to face alignment. However, due to the complex geometry of the facial appearance, the HG model still lacks the robustness of aligning faces in large poses. In this paper, a two-step method is proposed for robust face alignment. First, by using a convolutional neural network (CNN) to directly output the transformation parameters, the conventional procedure of normalizing the face region by performing Procrustes analysis based on the detected landmarks and the mean shape is simplified. In this way, faces with different poses can be converted to a canonical state, which is more advantageous for subsequent face alignment. Second, motivated by recent deformable convolutional networks, we propose a modulated deformable residual block and replace the plain counterparts in the HG model, resulting in deformable hourglass networks (DHNs). The DHN yields large performance improvements over original HG model while having the almost same amount of parameters and bringing minor additional computation costs. Depending on the synergistic effect of two innovations, the proposed method achieves better performance in comparison to the state-of-the-art methods on challenging benchmark datasets.
|