Joint Face Detection and Landmark Localization Based on an Extremely Lightweight Network.

Face detection and landmark localization are necessary steps in most face applications. At present, the method based on deep learning has shown obvious advantages in effect. However, most neural networks are computationally expensive and require special hardware for acceleration. To widely applied in real-world tasks, it is necessary to design a tiny model with fewer parameters, less computation cost, and fine performance. Therefore, we propose an extremely lightweight backbone for building a YOLOv3-style joint face detection and landmark localization model while compressing the parameters to the 0.15M level. We compare the proposed face detector with representative methods on the public benchmark. The results show that our proposed method can achieve performance much close to the representative face detector while a two-thirds reduction in the numbers of parameters and the computing costs. Moreover, our model has a lower failure rate (10%) of landmark localization and more robust.