Real time Face Mask Detector

This repository represents face mask detector, based on ultralytics YOLOv5s model and trained on custom dataset

Quick start

1. Install:

   git clone https://github.com/IvanKurnosov/mask_detector # clone repo
   cd mask_detector
   pip install -r requirements.txt # install requirements.txt

2. Run:

   python detect.py --source data/images/faces.jpg

This repo is a clone of ultralytisc/yolov5 with little changes. If you have it already installed, you may download weights/yolov5s_mask_detector_weights.pt and get a similar result by running:

python detect.py --weights weights/yolov5s_mask_detector_weights.pt --img 1024 --conf 0.5 --source data/images/faces.jpg

Training

The main problem in creating this detector was to generate good dataset. I found two face mask datasets: MedicalMaskDatasetImagesTfrecords and KaggleFaceMaskDetection. Both of them are not bad but they don't go into any comparison with WiderFace for example. So if you just train on this datasets, the model can't detect small or bad illuminated faces. After few experiments, I found two ways to ease this problems:

1. First is to pretrain model on WiderFace dataset and then train on both face mask dataset. It's improve results and you may use this model by running

    python detect.py --weights weights/yolov5s_mask_detector_wider_pretrained_weights.pt --source data/images/faces.jpg

2. Second is to 'wear' masks on 50% WiderFace datacet faces and use it together with mask datasets. It provides class balance and different lightnings, scales and positions. The disadvantage is that I use only 23 mask and they are weared on faces in similar way. You may find masks at data/images/masks.zip. This model is used by default.

Source code

All dataset generating and model training code stored in scripts/ folder:

1. WiderFace pretrained model:

wider_face_to_darknet.ipynb  # transform WiderFace dataset to Darknet format
wider_face_detector_train.ipynb  # train face detector
wider_pretrained_mask_detector train  # train mask detector

2. Custom WiderFace model:

custom_wider_kaggle_to_darknet.ipynb  # wear masks on WiderFace, union with Kaggle datasets and trainsform to Darknet format
mask_detector_custom_dataset_train.ipynb  # train mask detector

3. Both models use the same Yaml files:

faces.yaml  # to describe dataset
yolov5s.yaml  # to setup network