Multi-Level Feature Correction Inference Network for 3D Face Reconstruction and Dense Alignment

3D face reconstruction and dense alignment task has long been regarded as a challenging research problem, particularly in unconstrained environments, which has garnered significant attention in the computer vision community. In this paper, we propose a Multi-Level Feature Correction Inference Network (MFCINet), which is capable of recovering 3D faces with rich details and achieving more accurate dense alignment in unconstrained environments. Most traditional methods directly regress face identity and expression parameters, resulting in reconstructed 3D faces lacking fine-grained information. To address this issue, we utilize multi-level features to correct the features of the backbone network, yielding more robust prediction results. Specifically, we first employ a shallow network to learn the low-level spatial information of faces. This allows the model to better understand spatial relationships, such as relative positions, sizes, and orientations between facial components, by learning edge information and texture features. Secondly, we leverage a deep network to learn the high-level semantic information of faces, which enables the model to better capture semantic correlations in facial data and achieve a more accurate understanding and representation of facial semantic information. Finally, by combining the low-level spatial information and high-level semantic information, and applying a correction operation, we inject all the information into the backbone network for accurately reasoning. The effectiveness of our method is substantiated by extensive experiment conducted on the AFLW2000-3D and AFLW datasets.