AlexZaikin94 / Proper-Face-Mask-Object-Detection

Proper Face Mask Object Detection using PyTorch

Proper Face Mask Object Detection

a PyTorch implementation, using MobileNetV3 and SSDLite

Model Results

Using our trained model (CPU Inference)

the model was created with mobile deployment in mind, so it can be easily deployed on any machine with a webcam, following a few simple steps:

1. (optional, but recommended) create a clean Miniconda environment:

   conda create --name app python=3.8
   conda activate app
   

2. install requirements:

   pip install -r requirements_lite.txt
   

3. run the app:

   python app.py
   

4. (optional) the app can also be packed into a .exe file using pyinstaller:

   pip install pyinstaller
   pyinstaller --onefile -w app.py
   

   1. an antivirus exception may be needed for the .exe file.
   2. the .exe file should be ~270MB if using a clean Miniconda environment.

Re-Training the model

Enviorment

for training, use the environment.yml virtual environment:

conda env create -f environment.yml

Data

Training is done using 2 Kaggle datasets:

1. face-mask-detection by andrewmvd
2. face-mask-detection-dataset by wobotintelligence

data can be downloaded manually, or using the kaggle API:

1. install the kaggle package:

   pip install kaggle
   

2. create your Kaggle API key (in the Kaggle settings)

3. place your Kaggle API key .json file in ~/.kaggle/kaggle.json

4. download and extract the data:

   kaggle datasets download wobotintelligence/face-mask-detection-dataset
   unzip face-mask-detection-dataset.zip -d data/wobotintelligence
   rm face-mask-detection-dataset.zip
   
   kaggle datasets download andrewmvd/face-mask-detection
   unzip face-mask-detection.zip -d data/andrewmvd
   rm face-mask-detection.zip
   

Training (only single GPU supported)

training hyperparameters can be changed via the config.py file, for training run:

python main.py

or run main.ipynb if you preffer using a jupyter notebook.

Model

1. we use the torchvision implementation of the SSDLite-320x320 object detection model with a MobileNetV3-Large backbone: torchvision.models.detection.ssdlite320_mobilenet_v3_large.
2. we use a pretrained backbone, the object detection model is trained from scratch.