[AICUP 2023] Competition-2023-Pytorch-Badminton

🎉 This work has earned a place in the top 25% of winners!

Award winners list (Click to expand)

TEAM_2970: Jonathan, Joe, Dodo, Edward, Harry

• [AICUP 2023] Shuttlecock Recognization

Model architectures

Overall pipeline
MVD (Video Distillation)	DSNet (Video Summarization)	TrackNetv2 (Video Segmentation)

Each model purpose

• MVD (to predict the ball type)
• DSNet (to predict the shot frame number)
• TrackNetv2 (to predict the shuttlecock position, ball height)
• YOLOv8-Detection (to predict the player position)
• YOLOv8-PoseEstimation (to predict the toe position)
• YOLOv8-Classification (to predict the round head, backhand)

Installation

• Clone the code from repository

  git clone https://github.com/FanChiMao/Competition-2023-PyTorch-Badminton
  

• Install submodule

  cd Competition-2023-PyTorch-Badminton
  git submodule update --init
  

• Build the environment

  cd Competition-2023-PyTorch-Badminton
  pip install -r requirements.txt
  

• Download the TrackNetv2

  git clone https://nol.cs.nctu.edu.tw:234/open-source/TrackNetv2
  python utils\predict_process\TrackNet_import.py
  

Datasets

• Public Training data
• Public Testing data
• Private Testing data

Inference

• Download our YOLOv8 trained weights by following commands, or you can directly download from here.

  cd trained_weights
  python download_trained_weights.py
  

• Check the configuration path from ./inference.yaml

  # Path setting
  PATH:
    VIDEO: D:\AICUP\datasets\test\video
    HIT_CSV: D:\AICUP\datasets\test\predict_csv
    RESULT: .\predict_result
    OPENPOSE: .\for_openpose
  
  # Pretrained weights path
  WEIGHTS:
    PLAYER: D:\AICUP\Competition-2023-PyTorch-Badminton\trained_weights\yolov8s-players_detection_2.pt
    COURT: D:\AICUP\Competition-2023-PyTorch-Badminton\trained_weights\yolov8s-seg_net_detection.pt
    NET: D:\AICUP\Competition-2023-PyTorch-Badminton\trained_weights\yolov8s-players_detection.pt
    ROUNDHEAD: D:\AICUP\Competition-2023-PyTorch-Badminton\trained_weights\yolov8n-cls_roundhead.pt
    BACKHAND: D:\AICUP\Competition-2023-PyTorch-Badminton\trained_weights\yolov8n-cls_backhand.pt
    BALLTYPE: D:\AICUP\Competition-2023-PyTorch-Badminton\trained_weights\yolov8s-cls_balltypes.pt
    START: D:\AICUP\Competition-2023-PyTorch-Badminton\trained_weights\yolov8n-cls_balltypes_start.pt
    AFTER: D:\AICUP\Competition-2023-PyTorch-Badminton\trained_weights\yolov8n-cls_balltypes_after.pt
  

• After setting the configuration, to predict the video input, simply run

  python main_predict.py
  

Reference

• https://github.com/ultralytics/ultralytics
• https://github.com/ruiwang2021/mvd
• https://github.com/li-plus/DSNet
• https://nol.cs.nctu.edu.tw:234/open-source/TrackNetv2

Contact us

• Chi-Mao Fan: [email protected]
• Hong-Ru Shen: [email protected]