kosuke1701/docker-anime-face-detection

Dockerfile to detect faces and their landmark points from anime illustrations.

Getting Started

Prerequisite

Following softwares are required to build the docker image.

• nvidia-docker
  • https://github.com/NVIDIA/nvidia-docker
  • Edit docker configuration to change default runtime to nvidia.

    •   # E.g. edit /etc/docker/daemon.json as follows.
        {
            "default-runtime": "nvidia",
            "runtimes": {
                "nvidia": {
                    "path": "nvidia-container-runtime",
                    "runtimeArgs": []
                }
            }
        }
      

• nvidia driver compatible to CUDA 11.3

Building Docker Image

docker build -t face_pose docker

Start API

docker run --gpus all -p 32011:32011 -t face_pose uvicorn api:app --host 0.0.0.0 --port 32011

After API is started, you can access API as follows:

curl -X POST -F file=@/path/to/image http://localhost:32011/detect_faces

Returned json data is a list which contains a dictionary for each detected face with following keys. (OpenCV's coordinate system is used.)

• "bbox"
  • bounding box of a detected face
  • [xmin, ymin, xmax, ymax]
• "keypoints"
  • 28 landmarks points of a detected face (see https://github.com/hysts/anime-face-detector)
• "angle"
  • face rotation angle
• "valid_score"
  • If this value is low, it is more likely that detected landmark points are anomalous.
• "valid_score_threshold"
  • Default threshold for values of "valid_score".
• "features"
  • Normalized and flattened landmark positions for each face components.
• "origin", "scaleX", "scaleY"
  • Values used to transform landmark positions when computing values of "features".