AICITY2024_Track5

This repo includes solution for AICity2024 Challenge Track 5 - Detecting Violation of Helmet Rule for Motorcyclists

Installation

Please find installation instructions for PyTorch and YOLOv8 in here

Please find installation instructions for PyTorch and Co-DETR. Tested on torch1.11.1+cuda11.3+Pillow10.2.0 here

Data Preparation For Training

Download the training data (aicity2024_track5) into ./data/

The format of dataset as follows:

• data
  • aicity2024_track5
    • videos
    • ReadMe.txt
    • gt.txt
    • ...

cd data_pre

Extract frames and split train-val

python video2picture.py -f=300
python split_train_val.py
python yolo2coco.py

Generate label for YOLOv8

python original_9cls_to_bigbox.py

Generate label for Co-DETA

python create_crop_picture.py
python yolo2codetr.py

Train

Train YOLOv8(Coarse Detector)

cd stage1_coarse detector/ultralytics-main
python train.py

Train Co-DETR(Fine-grained Detector)

cd stage2_grained-fine detector/Co-DETR-main
CUDA_VISIBLE_DEVICES=0,1,2,3,4,5  bash tools/dist_train.sh projects/configs/my_model/my.py 6 /data1/zhp/2024AICITY/Co-detr/path

Inference

The checkpoints after trainning process can be downloaded here, which includes all the checkpoints. After downloading all the checkpoints, please put all files into ./weights/

cd data_pre

Extract frames

python video2picture.py
python yolo2coco.py

Detect results

cd stage1_coarse detector/ultralytics-main
CUDA_VISIBLE_DEVICES=6 yolo task=detect mode=val batch=8 workers=2 split=test save_json=True augment=True device=6 model=./weights/stage1_best.pt data=ultralytics/cfg/datasets/aicity.yaml 
name=stage1_coarse detector/ultralytics-main/run/stage1_result

cd ../../data_pre
python crop_predict_val_and_test.py

cd ../stage2_grained-fine detector/Co-DETR-main
CUDA_VISIBLE_DEVICES=0,1,2,3 bash tools/dist_test.sh projects/configs/my_model/my.py ./weights/stage1_epoch10.pth 4 --format-only --options "jsonfile_prefix=/data1/zhp/2024AICITY/Co-detr/inference_two_cls/test_epoch10"

cd ../../post_processing
python create_final_json.py

Public Leaderboard

TeamName	Score
BUPT_MCPRL	0.3940