FaceMask Detection

Introduction

Implementation of facemask detection with Yolor in PyTorch.

SetUp

• Download the dataset from Kaggle [https://www.kaggle.com/datasets/andrewmvd/face-mask-detection]
• Clone the Yolor from GitHub repo [https://github.com/WongKinYiu/yolor] into a folder FaceMask_Detection

Installation

• Install the requirements: pip install -r /yolor-main/requirements.txt

Required file changes

• Go to ..\FaceMask_Detection\yolor-main\data

  • Create the file custom.names:
    • Copy and paste the following classes:
      • without_mask
      • with_mask
      • mask_weared_incorrect
  • Create the yaml file custom.yaml:
    • Copy and paste the respective paths and classes:
      • train: ../data/train.txt
      • val: ../data/test.txt
      • names: ['without_mask', 'with_mask', 'mask_weared_incorrect']

• Go to ..\FaceMask_Detection\yolor-main\cfg

  • Create the config file yolor_p6_custom.cfg:
    • Change the number of classes to 3 → lines: 1614, 1658, 1702, 1746
    • Change the number of filters to 24 → [(number of classes + 5) * 5] → lines: 1605, 1649, 1693, 1737
    • Change the number of filters to 24 → [(number of classes + 5) * 5] → for: # 207, # 208, # 209, # 210 lines:1569, 1573,1577 and 1581 respectively. Keep everything unchanged.

Data Preparation

• Run process.py in ..\FaceMask_Detection\process.py to resize and save the images to 640 by 640
• Run covert.py in ..FaceMask_Detection\data\annotations\convert.py → Create text annotation files in Yolor format
• Copy the text annotation files to ..\FaceMask_Detection\data\images → The images and txt annotation files should be in the same folder

Training

• Run the train.py → To train the model
• python train.py --batch-size 2 --img 640 640 --data .\data\custom.yaml --cfg .\cfg\yolor_p6_custom.cfg --device 0 --name yolor_p6 --hyp .\data\hyp.scratch.640.yaml --epochs 300
• The batch size and the number of epochs can be changed as required

Prediction

• The best model saved: \yolor-main\runs\train\yolor_p6\weights\best.pt
• Run the detect.py → python .\detect.py --source .\inference\images\ --cfg .\cfg\yolor_p6_custom.cfg --weights .\yolor-main\runs\train\yolor_p6\weights\best.pt --conf 0.25 --img-size 640 --device 0 --names .\data\custom.names

Results

The results are as follows:

Citation

@misc{make ml, title={Mask Dataset}, url={https://makeml.app/datasets/mask}, journal={Make ML} }

@article{wang2021you, title={You Only Learn One Representation: Unified Network for Multiple Tasks}, author={Wang, Chien-Yao and Yeh, I-Hau and Liao, Hong-Yuan Mark}, journal={arXiv preprint arXiv:2105.04206}, year={2021} }