YOLOv7-Pose-on-Custom-Dataset:

Keypoint detection on custom dataset. We have 1 class - Glass and it have 4 keypoints. In this this tutorial we will train our yolov7 model to detect these 4 custom keypoints

Data Prepration:

Install Docker on your system if you haven't already. You can download it from the official website: https://www.docker.com/get-started

Now,Installing coco-annotator using docker:

git clone https://github.com/jsbroks/coco-annotator.git 

cd coco-annotator 

docker-compose up

http://localhost:5000/

Learn how to annotate: https://www.youtube.com/watch?v=OMJRcjnMMok&t=1s  

Next step is to convert json format annotations into YOLO format.

Refernce for conversion:  https://github.com/WongKinYiu/yolov7/issues/1103

Now, Dataset is ready. Clone this github repo:

    git clone https://github.com/AarohiSingla/YOLOv7-POSE-on-Custom-Dataset

    cd YOLOv7-POSE-on-Custom-Dataset

    pip install -r requirements.txt

    Place your dataset folder in this repo.

Make all the changes which are mentioned in this video: https://youtu.be/OP-oiDsEVzc

For Training:

  !python train.py --data data/custom_kpts.yaml --cfg cfg/yolov7-w6-pose_custom.yaml --hyp data/hyp.pose.yaml --device 0 --kpt-label --epochs 600

For Keypoint Detection:

 !python detect.py --weights runs/train/exp3/weights/best.pt --kpt-label --source 1.jpg --conf 0.030 --iou 0.30

Reference:

    https://github.com/WongKinYiu/yolov7/tree/pose

    https://github.com/ruiz-manuel/yolov7-pose-custom