/AICITY2023_Track5_DVHRM

Primary LanguagePython

AICITY2023_Track5_DVHRM

The solutions ranked fourth, fifth, and sixth in Track 5 (Detecting Violations of Helmet Rules for Motorcyclists) of the NVIDIA AI City Challenge at the CVPR 2023 Workshop.

Solution pipelines

  1. Download the training_videos from track5 of AI CIty Challenge
  2. Extracts images from Track 5 training_videos:
  • bash ./run_extract_train_frames.sh
  1. Convert gt.txt to yolo txt:
  • python GTxywh2yolo.py
  1. Uses YOLOv7-E6E to train the seven classes Helmet detector with 100 training videos and 100 validation videos:
  • python -m torch.distributed.launch --nproc_per_node 4 --master_port 9527 train_aux.py --workers 8 --device 0,1,2,3 --sync-bn --batch-size 8 --data Helmet/Helmet.yaml --img 1920 1920 --cfg cfg/training/yolov7-e6e-Helmet.yaml --weights '' --name yolov7-e6e-Helmet --hyp data/hyp.scratch.p6.yaml --epochs 350
  1. Uses YOLOv7-CBAM to train the seven classes Helmet detector with 100 training videos and 50 validation videos:
  • python -m torch.distributed.launch --nproc_per_node 4 --master_port 9527 train_aux.py --workers 8 --device 0,1,2,3 --sync-bn --batch-size 12 --data Helmet/Helmet.yaml --img 1280 1280 --cfg cfg/training/yolov7-e6e-CBAM-Helmet.yaml --weights '' --name yolov7-e6e-CBAM-Helmet1280-10050 --hyp data/hyp.scratch.p6.yaml --epochs 300
  1. Uses YOLOv7-SimAM to train the seven classes Helmet detector with 75 training videos and 25 validation videos:
  • python -m torch.distributed.launch --nproc_per_node 4 --master_port 9527 train_aux.py --workers 8 --device 0,1,2,3 --sync-bn --batch-size 12 --data Helmet/Helmet.yaml --img 1280 1280 --cfg cfg/training/yolov7-e6e-siam-Helmet.yaml --weights best.pt --name yolov7-e6e-siam-Helmet1280-7525 --hyp data/hyp.scratch.p6.yaml --epochs 300
  • Break at epoch 174
  1. Uses YOLOv7-SimAM to fine-tune the seven classes Helmet detector with 100 training videos and 50 validation videos:
  •  python -m torch.distributed.launch --nproc_per_node 4 --master_port 9527 train_aux.py --workers 8 --device 0,1,2,3 --sync-bn --batch-size 12 --data Helmet/Helmet.yaml --img 1280 1280 --cfg cfg/training/yolov7-e6e-siam-Helmet.yaml --weights best.pt --name yolov7-e6e-siam-Helmet1280-10050 --hyp data/hyp.scratch.p6.yaml --epochs 300
  1. Download the testing_videos from track5 of AI CIty Challenge
  2. Extracts images from Track 5 testing_videos:
  • bash ./run_extract_test_frames.sh
  1. Rank 6: Test YOLOv7-E6E seven classes Helmet detector
  • Download YOLOv7-E6E seven classes Helmet detector model and reanme it to best.pt
  • python detect_Helmet.py --source ../test --weights best.pt --conf 0.93 --iou-thres 0.45 --img-size 1920 --device 0 result_file_name y7e6eHelmet1920_93_45.txt
  1. Rank 5: Test YOLOv7-CBAM seven classes Helmet detector
  • Download YOLOv7-CBAM seven classes Helmet detector model and rename it to best.pt
  • python detect_Helmet.py --source ../test --weights best.pt --conf 0.93185 --iou-thres 0.35 --img-size 1280 --device 1 result_file_name y7e6eCBAMHelmet1280_93185_35.txt
  1. Rank 4: Test YOLOv7-SimAM seven classes Helmet detector
  • Download YOLOv7-SimAM seven classes Helmet detector model and rename it to best.pt
  • pthon detect_Helmet.py --source ../test_images --weights best.pt --conf 0.92 --iou-thres 0.35 --img-size 1280 --device 2 result_file_name y7e6esiamHelmet1280_92_35.txt

Environment

Please refer to YOLOv7 Installation