TF-Attn3DRecon: Transformer and Adversarial Attention Fusion for 3D Face Reconstruction
Zewang Zhang, Zhongjie Huang, Shengxiang Mei, Guang Li, Hua Wang Shi · 2025
3D face reconstruction technology is susceptible to constraints from viewing angles, illumination variations, and occlusions, leading to limited reconstruction quality. This study proposes an innovative method named TF-Attn3DRecon, featuring a dual-branch generator integrating CNN and Transformer, and a PatchGAN discriminator enhanced with channel attention (SE-Net). The generator synergizes local features with global contextual information to effectively decouple factors such as shape and facial expressions while enhancing detail capture capability. The PatchGAN discriminator incorporates patch-wise input and SE-Net to focus on critical facial regions and improve reconstruction robustness under occlusions. Evaluated on the FaceScape-wild (FaceScape-lab) dataset, the proposed method achieves Chamfer Distance (CD) and Mean Normal Error (MNE) metrics of 3.12 mm and 0.087 rad (3.55 mm and 0.101 rad), surpassing other comparative models. For side face and partially occluded face images, the visual reconstruction quality of our method outperforms existing models, particularly in texture reconstruction. Therefore, this method has a higher degree of face shape and texture restoration, and can maintain high integrity when reconstructing 3D faces.