ML Project: A face mask detector

Team member: Peng Tianyi, Wang Shuai, Ye Huanyi Based on 853 pictures and three classes (wear mask, no mask, wear mask incorrectly), we improved and trained the face mask detector based on YOLOv5s.

Performance

We get 80.9 mAP0.5 in the validation dataset with more than 100FPS in Tesla P100-PCIE-16GB.

The performance of our model

Class	#Labels	Precision	Recall	mAPval 0.5	mAPval 0.5:0.95
with mask	630	0.828	0.935	0.951	0.647
without mask	104	0.632	0.885	0.865	0.558
mask weared incorrectly	20	0.72	0.65	0.612	0.372
total	754	0.727	0.823	0.809	0.526

The improvement of our model

Prepare

Download data

• Down load data from my drive
• Unzip the file in root, replace datasets directory via unzip file
• Format like this

Environment

• We strongly recommend you to run in Google Colab or other Linux system.
• Install necessary part

pip install -r requirements.txt

Run!

Inference

• Open demo.ipynb
• Set path
• Run the cells

Validation

• Just continue running the cells

Train

• Please use GPU and run the cell
• If you want to freeze, please train with freeze firstly, then train without freeze based on former weights.

Effect show

• Wear mask, and wear mask incorrectly

* No mask

Reference

https://github.com/ultralytics/yolov5
https://github.com/spacewalk01/yolov5-face-mask-detection