Towards Accurate Facial Landmark Detection via Cascaded Transformers

Accurate facial landmarks are essential prerequisites for many tasks relatedto human faces. In this paper, an accurate facial landmark detector is proposedbased on cascaded transformers. We formulate facial landmark detection as acoordinate regression task such that the model can be trained end-to-end. Withself-attention in transformers, our model can inherently exploit the structuredrelationships between landmarks, which would benefit landmark detection underchallenging conditions such as large pose and occlusion. During cascadedrefinement, our model is able to extract the most relevant image featuresaround the target landmark for coordinate prediction, based on deformableattention mechanism, thus bringing more accurate alignment. In addition, wepropose a novel decoder that refines image features and landmark positionssimultaneously. With few parameter increasing, the detection performanceimproves further. Our model achieves new state-of-the-art performance onseveral standard facial landmark detection benchmarks, and shows goodgeneralization ability in cross-dataset evaluation.