trellixvulnteam / smokingcallingdetection_xhcq

Smoking and Calling Detection Challenge of HuaLu Cup 2020, Rank 9.

Smoking and Calling Detection

Info

• Smoking and Calling Detection Challenge of HuaLu Cup 2020, Rank 9.
  • Solved by Classfication.
  • 4Labels: Normal, Calling, Smoking, Calling&Smoking.
• Explored models including ResNeSt and BiT, choose ResNeSt in the end.
  • Pretrained on ImageNet and tranfered to this dataset.
  • Data Augment such as resize, ColorJitter,RandomHorizontalFlip, RandomRotation and so on.
  • Dropout for Overfitting.
  • Adam Optimizer, MultiStepLR.
• mAP of Semifinals on A Rank list : 97.55
• mAP of Semifinals on B Rank list : 89.95
• Final Contest: Rank 9.
• Members: J.-F. Meng, Z. Cheng, C.-W. Deng, SK49B of 5G&Multimedia Lab.

Usage (ResNeSt)

• Clone repo and cd into hualucup-ResNeSt

• Install Requirements.

• train

  python train.py 

• test

  python predict.py

• docker

  • download docker image with BaiduDisk, please contact to me and I will send you link and code.
  • Install nvidia-docker.
  • run docker image

    docker run --shm-size 32G -it hualucup-18310588600:dcw_update /bin/bash
    cd /root/hualucup
    # predict
    python predict.py
    # train
    python train.py 
    # ...