Today is April 25, 2021, COVID-19 has affected countries all over the world. Turning back a few day ago, India has recorded approximately 2000 death case per day, which once again alert us about how dangerous this disease are.
In this project, our main purpose is to build a detection system, that able to detect a person is either wearing a mask. The system based on some YOLO [1] version for object detection, they are:
- YOLOv3
- YOLOv3 fastest
- YOLOv5
By using the pre-defined models that were provided and supported by ultralytics. We will compare the result between these models, and implement a simple web application for run these model.
| Model | size (pixels) |
mAPval 0.5:0.95 |
mAPval 0.5 |
mAPtest 0.5:0.95 |
mAPtest 0.5 |
Speed V100 (ms) |
params (M) |
|
|---|---|---|---|---|---|---|---|---|
| YOLOv5s | 640 | 65.4 | 93.1 | 65.4 | 93.2 | 6.3 | 7.3 | |
| YOLOv5m | 640 | 66.5 | 93.9 | 66.7 | 93.7 | 7.9 | 21.4 | |
| YOLOv5l | 640 | 65.8 | 93.9 | 66.9 | 93.8 | 12.1 | 47.0 | |
| YOLOv5x | 640 | 66.5 | 93.5 | 67.3 | 94.0 | 20.7 | 87.7 | |
| YOLOv3 fastest | 640 | - | - | - | - | - | ||
| YOLOv3-tiny | 640 | 55.7 | 87.9 | 55.6 | 87.8 | 3.4 | 8.8 | |
| YOLOv3-SSP | 640 | - | - | - | - | - | 63.0 | |
| YOLOv3 | 640 | - | - | - | - | - | 61.9 |
We public our training result in wandb for if you want to dig deeper inside each model's training process, then make sure check out our project in W&B.
The dataset is composed of WIDER Face [2] and MAFA [3]. WIDER Face dataset contains 32,203 images with 393,703 normal faces which are refered as non masked face, MAFA contains 30,811 images with 35,806 masked faces.
Due to the limition of computational power, we using the dataset composed by AIZOOTech which only contains 7959 images in total, have been splited the dataset into 3 part: Train, Valid and Test; and converted them into YOLO format. You can find our dataset in Kaggle
Or by running Kaggle API:
kaggle datasets download -d nguyenmanhdung/facemaskyolo
We've implemented a simple Flask application for demonstrate our work where located in /deployment folder.
The quick demo is in the figure below, where we can see the yolov3 tiny model have a acceptable accuracy and a ablity of detecting multiple faces.
However, it's noticable some error that the model's made in some common scenarios in the video demonstration at Youtube:
To run the Flask application, direct to the \deployment folder, install all the requirements and run the following command:
$ pip install -r requirements.txt
$ python app.py
The output will be like shown below, thus, access to http://127.0.0.1:5000/ (or the port you have configured) to open the application(the browser should as for Webcam permission).
* Serving Flask app "app" (lazy loading)
* Environment: production
WARNING: This is a development server. Do not use it in a production deployment.
Use a production WSGI server instead.
* Debug mode: on
* Running on http://127.0.0.1:5000/ (Press CTRL+C to quit)
* Restarting with stat
* Debugger is active!
* Debugger PIN: 219-729-123
You might notice that we also support an builed application through Dockerfile, which you can find at
Note: We still in process of developing this deployment, if you have anytrouble, feel free to contact us.
Dung Manh Nguyen (me)
- Github: manhdung20112000
- Email: [email protected]
Hai Phuc Nguyen
- Github: HaiNguyen2903
Hoang Huy Nguyen
- Github: hhoanguet
[1] Joseph Redmon et al.You Only Look Once: Unified, Real-Time Object Detection. 2016.arXiv:1506.02640 [cs.CV].
[2] Shuo Yang et al. “WIDER FACE: A Face Detection Benchmark”. In:IEEE Conference onComputer Vision and Pattern Recognition (CVPR). 2016.
[3] Adnane Cabani et al. “MaskedFace-Net–A dataset of correctly/incorrectly masked faceimages in the context of COVID-19”. In:Smart Health19 (2021), p. 100144.