kokitakeishi / yolov5-face-mask-detection

Face Mask Detection using YOLOv5

Face Mask Detection

Face mask detection is an object detection task that detects whether people are wearing masks or not in videos. This repo includes a demo for building a face mask detector using YOLOv5 model.

Dataset

The model was trained on Face-Mask dataset which contains 853 images belonging to 3 classes, as well as their bounding boxes in the PASCAL VOC format. The classes are defined as follows:

• With mask
• Without mask
• Mask worn incorrectly

Setup

• Clone this repo and install YOLOv5:

git clone https://github.com/spacewalk01/face-mask-detection
cd face-mask-detection

# Install yolov5
git clone https://github.com/ultralytics/yolov5
cd yolov5
pip install -r requirements.txt

Training

• Download Face-Mask dataset from Kaggle and copy it into datasets folder.
• Execute the following command to automatically unzip and convert the data into the YOLO format and split it into train and valid sets. The split ratio was set to 80/20%.

cd ..
python prepare.py

• Start training:

cd yolov5
python train.py --img 640 --batch 16 --epochs 100 --data ../mask_config.yaml --weights yolov5s.pt --workers 0

Inference

• If you train your own model, use the following command for inference:

python detect.py --source ../datasets/input.mp4 --weights runs/train/exp/weights/best.pt --conf 0.2

• Or you can use the pretrained model models/mask_yolov5s.pt for inference as follows:

python detect.py --source ../datasets/input.mp4 --weights ../models/mask_yolov5.pt --conf 0.2

Results

The following charts were obtained after training YOLOv5s with input size 640x640 on the Face Mask dataset for 100 epochs.

Precision-Recall (PR) measures the accuracy of bounding box overlap between predicted bboxes and ground truth bboxes.

Precision Curve	Recall Curve	PR Curve

The following metrics were measured on the valid dataset containing 171 images. High precision and recall scores were obtained on with mask and without mask classes. However, the performance on mask weared incorrectly class was poor due to the imbalanced data.

Class	#Labels	Precision	Recall	mAPval 0.5	mAPval 0.5:0.95
with mask	630	0.94	0.9	0.95	0.64
without mask	104	0.86	0.82	0.9	0.6
mask weared incorrectly	20	0.72	0.3	0.43	0.24
total	754	0.84	0.67	0.76	0.49

Ground Truths vs Predictions

Ground Truth	Prediction

Reference

• Darknet
• YOLOv5