Skip to content

Yolov7-pose with variable keypoint support. Trained models with COCO Wholebody.

License

Notifications You must be signed in to change notification settings

fan23j/yolov7-pose-whole-body

Repository files navigation

yolov7-pose-whole-body

Implementation of "YOLOv7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors" combined with "Whole-Body Human Pose Estimation in the Wild".

This repo seeks to combine the aforementioned papers/repos to add extra keypoints to yolo-pose models.

Pose estimation implimentation is based on YOLO-Pose.

Pretrained models

yolov7-tiny-pose

python train.py --data data/coco_kpts.yaml --cfg cfg/yolov7-tiny-pose.yaml --batch-size 64 --img 640 --kpt-label --sync-bn --device 0  --hyp data/hyp.pose.yaml --nkpt 133 --weights PATH_TO_PRETRAINED_WEIGHTS epochs 500

Dataset preparation

[Keypoints Labels of MS COCO 2017]

COCO Whole-Body: https://github.com/jin-s13/COCO-WholeBody

Handy COCO to YOLO conversion script in utils/coco2yolo.py.

Training

yolov7-w6-person.pt

python -m torch.distributed.launch --nproc_per_node 8 --master_port 9527 train.py --data data/coco_kpts.yaml --cfg cfg/yolov7-w6-pose.yaml --weights weights/yolov7-w6-person.pt --batch-size 128 --img 960 --kpt-label --sync-bn --device 0,1,2,3,4,5,6,7 --name yolov7-w6-pose --hyp data/hyp.pose.yaml

Deploy

TensorRT:https://github.com/nanmi/yolov7-pose

Testing

yolov7-w6-pose.pt

python test.py --data data/coco_kpts.yaml --img 960 --conf 0.001 --iou 0.65 --weights yolov7-w6-pose.pt --kpt-label

Citation

@article{wang2022yolov7,
  title={{YOLOv7}: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors},
  author={Wang, Chien-Yao and Bochkovskiy, Alexey and Liao, Hong-Yuan Mark},
  journal={arXiv preprint arXiv:2207.02696},
  year={2022}
}

Acknowledgements

Expand

About

Yolov7-pose with variable keypoint support. Trained models with COCO Wholebody.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published