Pixel-in-Pixel Net: Towards Efficient Facial Landmark Detection in the Wild

Recently, heatmap regression models have become popular due to their superior performance in locating facial landmarks. However, three major problems still exist among these models: (1) they are computationally expensive; (2) usually lack explicit constraints on global shapes; (3) domain gaps commonly present. To address problems, we propose Pixel-in-Pixel Net (PIPNet) for landmark detection. The proposed model is equipped with a novel detection head based regression, which conducts score and offset predictions simultaneously low-resolution feature maps. By doing so, repeated upsampling layers no longer necessary, enabling the inference time be largely reduced without sacrificing accuracy. Besides, simple but effective neighbor module enforce local by fusing from neighboring landmarks, enhances robustness of new head. further improve cross-domain generalization capability PIPNet, self-training curriculum. This training strategy able mine more reliable pseudo-labels unlabeled data across domains starting an easier task, then gradually increasing difficulty provide precise labels. Extensive experiments demonstrate superiority obtains state-of-the-art results out six benchmarks under supervised setting. two test sets also consistently improved compared baselines. Notably, our lightweight version PIPNet runs at 35.7 FPS 200 CPU GPU, respectively, while maintaining competitive accuracy methods. code available https://github.com/jhb86253817/PIPNet.