Face Cropper

Stripped down functionality from Face Crop Plus using AMD's RetinaFace ONNX with no PyTorch dependancy, a 1.7 MB model artifact and a single file of code for inference.

Why

What I need: square image of one face from image of one face that is not square.

What I don't need: 7GB venv with CUDA PyTorch dependencies.

How

Detects bbox of a face (one face) in the picture (using RetinaFace and ONNX for inference) then crops a square around it with the face being in the center (cv2 slicing).

Setup

Local

uv venv
uv pip install -r requirements.txt
. .venv/bin/activate

Docker (API)

Build

docker build -t cropper .
docker run -d -p 8080:8080 cropper:latest

Pull

docker pull ghcr.io/a-tabaza/center_crop:main
docker run -d -p 8080:8080 ghcr.io/a-tabaza/center_crop:main

Go to localhost:8080/docs for SwaggerUI

Usage

CLI

python src/crop.py -i input.jpg -o output.jpg

API

Multiple images (result in .zip file)

curl -X 'POST' \
  'http://localhost:8080/crop/' \
  -H 'accept: image/jpeg' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@4.jpg;type=image/jpeg' \
  -F 'files=@3.jpeg;type=image/jpeg' \
  -F 'files=@2.jpg;type=image/jpeg' \
  -F 'files=@1.jpg;type=image/jpeg'

One image (result back as jpg)

curl -X 'POST' \
  'http://localhost:8080/crop/' \
  -H 'accept: image/jpeg' \
  -H 'Content-Type: multipart/form-data' \
  -F 'files=@1.jpg;type=image/jpeg' \

Examples

Contributing

No. This isn't for you, it's to save my time, if it helps, I'm glad, if it doesn't, fork and adapt it.