cpp-onnx-port

# cpp-onnx-port Use this sub-skill when the task is about the optional 3DDFA C++ demo, OpenCV DNN execution, MobileNet checkpoint export to ONNX, C++ weight placement, CMake/OpenCV build behavior, or C++ output failures. Do not use this sub-skill as the primary guide for Python inference, rendering internals, training, or benchmark evaluation. Route those tasks to sibling skills when available: - Python image/video inference and checkpoint choice: `../python-inference/SKILL.md` - Geometry, vertices, OBJ/PLY, PNCC, depth, and rendering internals: `../geometry-rendering/SKILL.md` - Training recipes, losses, checkpoints, and benchmarks: `../training-evaluation/SKILL.md` ## Operating Map 1. For C++ build and demo behavior, read `references/cpp-build-and-runtime.md`. 2. For checkpoint export, read `references/onnx-export.md` and use `scripts/export_mobilenet_to_onnx.py` instead of relying on the original conversion helper. 3. For missing faces, missing weights, OpenCV DNN, ONNX, CMake, or output-image failures, read `references/troubleshooting.md`. ## Key Constraints - The C++ port is optional and unoptimized; it relies on OpenCV DNN and was documented for OpenCV 4.2.0 or newer. - The C++ demo uses CPU OpenCV DNN backends for both the YOLO face detector and MobileNet landmark predictor. - The source distribution includes C++ matrix/config text files, but the MobileNet ONNX file and YOLO binary weights are external artifacts that must be supplied by the operator. - The C++ demo is hard-coded around a single sample image and writes a landmark-overlay image under its result directory unless the C++ source is adapted. - Keep `num_classes=62` for C++ landmark compatibility: 12 pose/projection values, 40 shape coefficients, and 10 expression coefficients. ## Safe Actions - Build only after confirming OpenCV development headers/libraries are available and the required external model files are present. - Export ONNX from a known 3DDFA MobileNet checkpoint with the bundled helper and inspect remapped checkpoint keys before trusting the file. - Treat CMake, OpenCV, detector, predictor, and output-path problems as separate failure classes; do not debug them as one generic C++ failure.