Adaptive Multi-Dimensional Spatial Mapping for 3D Face Reconstruction

As the application of 3D faces becomes increasingly widespread across various fields, reconstructing 3D faces with precise facial shapes and faithful emotional conveyance is of great significance. Recent progress in methods based on 3D Morphable Models (3DMMs), which regress parametric 3D facial models from single facial images, have proven effective in restoring the overall facial shape. However, these methods struggle to capture subtle, due to flaws in the supervision cues. This paper proposes the Adaptive Multi-Dimensional Spatial Mapping Network (AMDSM-Net) as a solution. By incorporating an adaptive photometric loss function, we effectively reduce the domain gap, thus improving the fidelity of 3D face reconstruction. In particular, we designed a joint learning paradigm which leverages a spatial relationship weighted photometric loss among 2D image and 3D mesh. Our qualitative and quantitative evaluation evaluation outcomes on the dataset verify the superior performance of the method.