---

name: 3ddfa description: "Guide 3DDFA Python inference, geometry rendering, training/evaluation, and optional C++ ONNX workflows for 3D dense face alignment." disable-model-invocation: true metadata: disco-role: operating license: MIT

3DDFA Repo Skill

Use this repo skill when a task involves 3DDFA / 3D Dense Face Alignment: face alignment in full pose range, 68-point landmark prediction, dense 3D face vertices, pose boxes, PLY/OBJ export, depth/PNCC/PAF outputs, MobileNet-V1 checkpoints, 3DDFA training/evaluation, or the optional C++ OpenCV DNN port.

This skill is an operating guide for a 3DDFA checkout or adapted codebase. It is self-contained: use the references and bundled scripts here for routing, command construction, diagnostics, and troubleshooting instead of reopening the original repository documentation.

First Checks

1. Read references/repo-provenance.md before deciding whether this skill matches a checkout.
2. Read references/install-and-compatibility.md before installing dependencies or choosing CPU/CUDA/dlib/Cython paths.
3. Run scripts/check_3ddfa_environment.py against the target checkout for a safe import/resource diagnostic.
4. If the task names a concrete workflow, route to the matching sub-skill below.

Minimal diagnostic from this skill root:

python scripts/check_3ddfa_environment.py --repo-root /path/to/3DDFA

The diagnostic checks resources and imports; it does not run native inference, training, downloads, CMake builds, or benchmarks.

Route Map

User task or signal	Read
Run still-image inference, no-dlib bbox inference, inspect main.py flags, diagnose dlib/Cython startup, verify MobileNet forward shape, understand output filenames	sub-skills/python-inference/SKILL.md
Decode 62-D parameters, ROI boxes, sparse/dense vertices, PLY/OBJ/.mat, pose matrices, depth/PNCC/PAF, Cython renderer, BFM/3DMM data artifacts, video-frame rendering	sub-skills/geometry-rendering/SKILL.md
Adapt training commands, choose WPDC/VDC/PDC, validate filelists/param files/data roots, resume checkpoints, interpret AFLW/AFLW2000 metrics	sub-skills/training-evaluation/SKILL.md
Export MobileNet checkpoint to ONNX, place C++ weights, build/run OpenCV DNN demo, debug CMake/OpenCV/Yolo/ONNX issues	sub-skills/cpp-onnx-port/SKILL.md
Cross-cutting install/import/runtime failure	references/troubleshooting.md

Operating Boundaries

• Prefer CPU-safe diagnostics first. CUDA training/evaluation and GPU inference are optional capability paths and must be verified separately.
• The unmodified Python image CLI imports dlib and render utilities before argument parsing. Even bbox-only workflows can fail at startup if Python dlib or the Cython render extension is missing.
• Depth and PNCC require the compiled Cython mesh core; PLY/OBJ/landmarks can be planned separately, but the native CLI import path may still require the extension unless wrapped or patched.
• Full training, benchmark extraction, and the C++ demo depend on external datasets, optional weights, system packages, or GPUs. Treat these as explicit prerequisites, not default verification steps.
• Do not use this skill for 3DDFA_V2 unless the user explicitly asks to port concepts; this skill is based on the legacy 3DDFA repository snapshot in the provenance reference.

Bundled Scripts

• scripts/check_3ddfa_environment.py — shared checkout/resource/import diagnostic.
• sub-skills/python-inference/scripts/inspect_3ddfa_inference.py — image/video inference-specific diagnostic and command planner.
• sub-skills/python-inference/scripts/smoke_mobilenet_forward.py — safe MobileNet architecture forward smoke.
• sub-skills/geometry-rendering/scripts/smoke_geometry.py — safe 3DMM reconstruction shape smoke.
• sub-skills/training-evaluation/scripts/validate_training_args.py — training command/data-layout checker that does not launch training.
• sub-skills/cpp-onnx-port/scripts/export_mobilenet_to_onnx.py — explicit ONNX export helper for the optional C++ port.

Verification Expectations

Safe verification usually includes:

• package/import/resource diagnostics;
• MobileNet CPU forward shape (1, 62);
• geometry reconstruction shapes (3, 68) and dense (3, 53215);
• script --help checks for bundled helpers;
• explicit skip notes for dlib predictor, Cython build, CUDA, external datasets, and OpenCV C++ demo when unavailable.

Do not claim native end-to-end inference, GPU training/evaluation, full benchmarks, or C++ runtime success unless those exact paths were run in the target environment.