3D Morphable Models (3DMMs) provide promising 3D face reconstructions invarious applications. However, existing methods struggle to reconstruct faceswith extreme expressions due to deficiencies in supervisory signals, such assparse or inaccurate landmarks. Segmentation information contains effectivegeometric contexts for face reconstruction. Certain attempts intuitively dependon differentiable renderers to compare the rendered silhouettes ofreconstruction with segmentation, which is prone to issues like local optimaand gradient instability. In this paper, we fully utilize the facial partsegmentation geometry by introducing Part Re-projection Distance Loss (PRDL).Specifically, PRDL transforms facial part segmentation into 2D points andre-projects the reconstruction onto the image plane. Subsequently, byintroducing grid anchors and computing different statistical distances fromthese anchors to the point sets, PRDL establishes geometry descriptors tooptimize the distribution of the point sets for face reconstruction. PRDLexhibits a clear gradient compared to the renderer-based methods and presentsstate-of-the-art reconstruction performance in extensive quantitative andqualitative experiments. Our project is available athttps://github.com/wang-zidu/3DDFA-V3 .