3DDFA_V2

Why we included this project

This project turns a single photo into a full 3D face model, estimating head pose and dense landmarks in one pass. It regresses 3DMM parameters through a lightweight MobileNet backbone and reconstructs over 30,000 points per face, with output options ranging from depth maps and PNCC to UV texture or .ply/.obj meshes. The repo includes ready-to-run demos for stills, video, and webcam, and the onnxruntime path brings CPU inference down to roughly 1-2ms per face, which is what makes real-time tracking and embedded use feasible. Teams building face filters, avatars, or AR effects can adapt the pipeline quickly using the pretrained weights and Colab tutorial. It is a research codebase rather than a maintained service, so expect to handle the Cython build step and some rough edges yourself.