When 68 face landmarks aren't enough

What it does

The repository hosts the official implementation and dataset for a CVPR 2022 paper on 3D head alignment. It provides training code, a benchmark evaluation protocol, and a demo that infers 68, 191, or 445 2D facial landmarks, dense face and head meshes, head pose, and FLAME parameters from a single image. The underlying dataset, now on Hugging Face, includes vertex-level annotations, projection matrices, and metadata tags for quality, occlusion, expression, and pose.

The interesting bit

Most face datasets settle for a handful of landmarks; this one annotates hundreds of points and full head geometry, then packages it with a ready-to-run demo that exports standard .obj meshes and FLAME parameters. The recent addition of a UV texture predictor and PNCC estimator suggests the maintainers are still expanding what a single photograph can yield.

Key highlights

• Dataset annotations include vertices, model-view matrices, and projection matrices alongside metadata for quality, occlusion, pose, expression, age, and gender.
• Demo supports multiple output modes: 68/191/445 2D landmarks, face and head meshes, head pose visualization, 3D .obj export, and FLAME parameter json.
• Training pipeline uses Hydra for configuration management.
• Released under CC BY-NC-SA 4.0—free for research, off-limits for commercial use.
• Dataset is available via Hugging Face; benchmark evaluation code lives in a dedicated subdirectory.

Caveats

• The code is strictly Python 3.8, which may annoy anyone running a modern environment.
• The non-commercial license means you cannot drop this into a paid product without negotiating terms.
• The README is silent on hardware requirements and inference speed, so performance characteristics are unclear.

Verdict

Worth a look if you are doing 3D face reconstruction, head-pose estimation, or need densely annotated head data for research. Skip it if you need a commercial-friendly license or a plug-and-play API without dataset wrangling.