Abstract
Accurately localizing landmarks on 3D faces is critical for various applications, such as expression recognition, facial surgery navigation, and lip shape analysis. Existing landmarks localization methods generally contain complex calculation processes, which may affecting the efficiency. To address this problem, we propose a Simple and Efficient MLP-based Network (SEMNet) for landmarks localization. We first design a lightweight enhanced geometric affine module to adaptively transform point features in local region, for improving performance and generalization. Then, to fully utilize the rotation information of the face, a rotation constraint auxiliary branch is introduced for assisting in locating landmarks. In addition, to generate more accurate results, we propose a residual graph convolution discriminator to distinguish predicted locations from real face point clouds locations. Experimental results on two public datasets (FRGC v2 and Bosphorus) show that our method achieves high accuracy and efficiency compared to state-of-the-art methods. The project page is available at https://github.com/leimingyangde/SEMNet.