YOLOv7 Face Detection

Custom Yolov7 to detect face and estimate landmark.

Install

• pip3 install -r requirements.txt
• pip3 install Cython
• cd evaluation && sudo python3 setup.py install
• daquexian/onnx-simplifier
• onnx-graphsurgeon
• If export tensorRT hiennguyen9874/TensorRT

Performance

Data preparation

• Download and extract winderface dataset.
• Download and extract annotation file retinaface_gt_v1.1.zip.
• Download ground_truth.
• Folder after download and extract all:

  - ./winderface
    - WIDER_test/
        - images/
            - 0--Parade/
            - ...
    - WIDER_train/
        - images/
            - 0--Parade/
            - ...
    - WIDER_val/
        - images/
            - 0--Parade/
            - ...
    - train/
        - labels.txt
    - val/
        - labels.txt
    - test/
        - labels.txt
    - ground_truth/
        - wider_easy_val.mat
        - wider_medium_val.mat
        - wider_hard_val.mat
        - wider_face_val.mat

• Convert label to yolo format:
  • python3 scripts/convert_to_yolo.py --root ./widerface --image-folder WIDER_train/images --label-file train/label.txt --txt-file train.txt
  • python3 scripts/convert_to_yolo.py --root ./widerface --image-folder WIDER_val/images --label-file val/label.txt --txt-file val.txt
• Modify path of winderface folder in data/winderface.yaml

Detect

• python3 detect.py --weights ./weights/yolov7-tiny.pt --source inference/images --img-size 640 --conf-thres 0.2 --iou-thres 0.5 --device 1 --no-trace

Testing

• python3 eval.py --weights ./weights/yolov7-tiny.pt --data-root ./winderface --img-size 640 --conf-thres 0.01 --iou-thres 0.5 --device 0 --no-trace

• python3 evaluation/main.py -p ./outputs -g ./winderface/ground_truth

Training

• Download file yolov7-tiny.pt and save as ./weights/yolov7-tiny-origin.pt.

• Single GPU training: python3 ./train.py --device 0 --batch-size 16 --data data/widerface.yaml --img 640 640 --cfg cfg/yolov7-tiny-landmark.yaml --weights ./weights/yolov7-tiny-origin.pt --name yolov7-tiny --hyp data/hyp.scratch.tiny.yaml --noautoanchor --linear-lr --epochs 80

• Multiple GPU training: torchrun --standalone --nnodes=1 --nproc_per_node 2 ./train.py --device 0,1 --batch-size 16 --data data/widerface.yaml --img 640 640 --cfg cfg/yolov7-tiny-landmark.yaml --weights ./weights/yolov7-tiny-origin.pt --name yolov7-tiny --hyp data/hyp.scratch.tiny.yaml --noautoanchor --sync-bn --linear-lr --epochs 80

Export

ONNX

• python3 export.py --weights ./weights/yolov7-tiny.pt --img-size 640 --batch-size 1 --dynamic-batch --grid --end2end --max-wh 640 --topk-all 100 --iou-thres 0.5 --conf-thres 0.2 --device 1 --simplify --cleanup

• tools/Yolov7onnx_mask.ipynb

TensorRT

• Install custom TensorRT plugin

• python3 export.py --weights ./weights/yolov7-tiny.pt --img-size 640 --batch-size 1 --dynamic-batch --grid --end2end --max-wh 640 --topk-all 100 --iou-thres 0.5 --conf-thres 0.2 --device 1 --simplify --cleanup --trt

• /usr/src/tensorrt/bin/trtexec --onnx=./weights/yolov7-tiny.onnx --saveEngine=./weights/yolov7-tiny-nms-trt.trt --workspace=8192 --fp16 --minShapes=images:1x3x640x640 --optShapes=images:1x3x640x640 --maxShapes=images:4x3x640x640 --shapes=images:1x3x640x640

• tools/YOLOv7trt_mask.ipynb

Bugs

• val/llmks_loss = 0?

TODO

• How to achieve mAP of yolov7-tiny in deepcam-cn/yolov5-face

Acknowledgments

• WongKinYiu/yolov7
• deepcam-cn/yolov5-face