An Efficient Anchor-Based Face Alignment Network with Transformer

Despite significant advances have been made in facial alignment recently, face alignment remains a challenging problem due to the existence of issues like occlusion and large pose. Besides, small attention has been paid to the algorithm's performance, efficient face landmark localization algorithm with high robustness still has room to enhance. In this work, we propose an efficient face alignment network that combines the Transformer with an anchor-based prediction method. First, we extract features of the input image by CNNs, then capture long-range relationships efficiently using Transformer encoders, at last, we use anchor points to predict landmark coordinates. We test our algorithm through experiments on WFLW, the popular face alignment benchmark. The experiments show that our algorithm can reach high accuracy with satisfactory robustness while also enjoying the high speed.