Refined dense face alignment through image matching

Face alignment is the foundation of building 3D avatars for virtue communication in the metaverse, human-computer interaction, AI-generated content, etc., and therefore, it is critical that face deformation is reflected precisely to better convey expression, pose and identity. However, misalignment exists in the currently best methods that fit a face model to a target image and can be easily captured by human perception, thus degrading the reconstruction quality. The main reason is that the widely used metrics for training, including the landmark re-projection loss, pixel-wise loss and perception-level loss, are insufficient to address the misalignment and suffer from ambiguity and local minimums. To address misalignment, we propose an image MAtchinG-driveN dEnse geomeTrIC supervision (MAGNETIC). Specifically, we treat face alignment as a matching problem and establish pixel-wise correspondences between the target and rendered images. Then reconstructed facial points are guided towards their corresponding points on the target image, thus improving reconstruction. Synthesized image pairs are mixed up with face outliers to simulate the target and rendered images with ground-truth pixel-wise correspondences to enable the training of a robust prediction network. Compared with existing methods that turn to 3D scans for dense geometric supervision, our method reaches comparable shape reconstruction results with much lower effort. Experimental results on the NoW testset show that we reach the state-of-the-art among all self-supervised methods and even outperform methods using photo-realistic images. We also achieve comparable results with the state-of-the-art on the benchmark of Feng et al. Codes will be available at: github.com/ChunLLee/ReconstructionFromMatching.