OpenPose Docker Runner

This is the Docker path for HAND-005. It keeps OpenPose out of the host Python/Node environment and gives the hand-comparison pipeline a reproducible binary target.

Build

cd fronts/openpose
docker build -t labis/openpose-hand:1.7.0 -f docker/Dockerfile .

The build is heavy because OpenPose compiles Caffe/CUDA and downloads the official BODY_25 and hand models.

Run On A Folder

cd fronts/openpose
mkdir -p ../manos-agarres/results/comparison/openpose
docker run --rm --gpus all \
  -v "$PWD/../manos-agarres/data/synthetic:/data/input:ro" \
  -v "$PWD/../manos-agarres/results/comparison/openpose:/data/output" \
  labis/openpose-hand:1.7.0 \
  --image_dir /data/input \
  --hand \
  --write_json /data/output \
  --display 0 \
  --render_pose 0

CPU fallback is possible by removing --gpus all, but OpenPose hand inference is slow and may be impractical for CI.

Convert Outputs

OpenPose writes per-image JSON files. Normalize them into the common schema with:

cd fronts/manos-agarres
python scripts/normalize_hand_output.py --openpose-dir results/comparison/openpose

License Gate

OpenPose is free for non-commercial use. Before using it in a product or public service, confirm the commercial-license path from the upstream CMU OpenPose project.