Face alignment combined with shape constraints and Gaussian heatmap

Face alignment is a hot topic in the field of computer vision because it is an important intermediate step in face recognition, 3D face reconstruction, etc. Its task is to locate the key parts of the human face with landmarks. For the lack of shape constraints and weak correlation between landmarks in face alignment based on Gaussian heatmap regression, this paper presents a novel method combining shape constraints with Gaussian heatmap regression network. Our method consists of a 3D model parameter prediction network, an attention map generator, and a Gaussian heatmap regression network. The 3D model parameter prediction network fits the 3D face model and sample face. The attention map generator can generate an attention map containing shape constraint information based on the fitting results. The attention map and the original image will be used together as the input of the Gaussian heatmap regression network to achieve coarse to fine face alignment. The experimental results demonstrate that our method has a lower error on multiple mainstream datasets and is robust for some complex samples compared with the Gaussian heatmap regression method without shape constraints.