MDFN: Multi-path Dynamic Fusion Network for Face Reconstruction and Dense Face Alignment
Kangbo Wu, Zhiyuan Zhou, Xiaoxiao Yang, Xueming Wang · 2021 China Automation Congress (CAC) · 2021
3D face reconstruction based on single-view in the wild has been a long-standing challenging problem. In the complex and changeable unconstrained state, the traditional 3D Morphable Model (3DMM) parameters regression method lacks the ability to express local details, which is hard to reconstruct the accurate face shape. In this paper, we propose a multi-path pyramidal convolution network with the dynamic fusion of global and local information, which is able to recover richer detail 3D shapes from 2D images. Specifically, we design a multi-path network to extract global discriminative features and local detailed features, respectively. Then, we utilize the attention transformer module to enhance the network’s ability for capturing the correlation between local information. Finally, to boost the regression accuracy of the network, we dynamically fuse the global information that constrains the geometric shape of the face and the local information that enriches geometric details. Extensive experimental results on the AFLW and AFLW2000-3D datasets demonstrate that our MDFN achieves compelling performance in dense face alignment and reconstruction.