alphapose-feedback/README.md

Looping Alphapose Training

Pulling the traditional media art trick: can we loop the training of Alphapose? What if we train it on COCO data, run it on COCO images, train it on these results, run it on COCO images etc. ad inf.?

Installation

poetry install
git submodule init
git submodule update

Then make sure alphapose-docker is set up:

cd alphapose-docker
docker build --tag alphapose .

As mentioned in the README from alphapose-docker it is necessary to download some external models before usage, as per the AlphaPose installation guide.

1. YOLOv3 Object detector can be place in the alphapose-docker/detector/yolo/data repository.
2. (Optionally) YOLOX models go in the alphapose-docker/detector/yolox/data repository.
3. A pretrained AlphaPose model can be place in the alphapose-docker/pretrained_models directory. See their Model Zoo for the various options.
4. For pose tracking, see the pose tracking module. Make sure to add the necessary folder as a volume to the docker run command.

Then make sure to download COCO keypoint 2017 to the data directory:

./downloadCOCO.sh

Usage

Then the scripts can be invoked with poetry run python ....py