Full-pose 3D face reconstruction from a single snapshot

What it does

3DDFA is a PyTorch reimplementation of a 2017 face-alignment paper that estimates dense 3D facial geometry—including 68 landmarks, vertex-level shape, pose, and depth—from a single unconstrained image. It ships with pre-trained MobileNet-V1 models and can export textured .obj meshes, point clouds, and auxiliary feature maps like PNCC and PAF.

The interesting bit

The project goes beyond the original paper by adding real-time training strategies, a Cython-accelerated render pipeline, and a simple C++ port. The author also released a successor (3DDFA_V2), so this repo sits in an odd spot: it is the improved PyTorch version of the 2017 work, yet no longer the latest iteration.

Key highlights

• Single-image inference outputs 3D meshes, pose angles, depth maps, and PNCC/PAF features simultaneously.
• GPU inference clocks in at roughly 0.27 ms per image when batching 128 frames on a TITAN X; CPU inference on a modern MacBook Pro lands around 14.5 ms.
• Built on MobileNet-V1, making the model compact enough to consider for mobile deployment.
• Includes full training scripts, loss variants (WPDC, VDC, PDC), and pre-processed dataset configurations.
• Optional dlib integration for face detection, but supports landmark-free cropping and custom bounding boxes.

Caveats

• The repo omits large binary files such as the dlib detection model, so you must download them separately before running anything.
• Windows is explicitly untested; Linux and macOS are the supported platforms.
• The README openly points users to 3DDFA_V2 (ECCV 2020) for the latest pre-trained models and code, so this version appears to be in spare-time maintenance.

Verdict

Good for researchers or hackers who need a fully open, training-ready 3D face pipeline in PyTorch and do not mind using slightly dated weights. Skip it if you want the latest accuracy or a turnkey, cross-platform tool.