Face Mask Detection (You-Only-Look-Once v8)

Here's the link to the Google Colab notebook.

Installation

1. Create a new environment

python -m venv .venv
.venv/Scripts/activate # on windows
or
source .venv/bin/activate # on linux

2. Create a new .env file and edit file paths.

DATA_DIRECTORY_PATH = 'data/'   # path to root dataset directory
ANNOTATION_PATH = 'data/cleaned_data.csv'   # path to cleaned csv containing bboxes
IMAGES_OUTPUT_DIRECTORY_PATH = 'data/images'    # path to selected images from the entire dataset
CLEANED_IMAGES_PATH = 'data/images'     # path to cleaned images directory
TRAINING_CLS_DATA = 'data/data_cls'     # path to training classification data

# Set your paths here
FILE_PATH = 'path/to/your/CSV'
IMAGES_DIR ='path/to/your/images/directory'
VIDEOS_DIR = 'path/to/your/video/directory'

3. Install Dependencies

pip install -r requirements.txt

4. Run app for training

python main.py  

or perform inference from the cli using,

yolo predict model='models\best.pt' source=' .mp4/.jpeg/etc'

Summary

• Object Detection - Trained on YOLOv8n.pt (Model getting better on higher epochs)
• Classification - Trained on YOLOv8n-cls.pt
• Mask, No Mask; Detection and Classification.
• Custom Dataset (using CV). Labels corrected using OpenCV and observation using preview.py.
• Trained on YOLOv8, locally for both classification and object-detection.
• Classes 0: Mask, 1: No-Mask
• Checkpoint for classification - 'models/best.pt', object detection - 'models/best_od.pt'
• Helper functions for both classification and object detection in utils.
• Output stored at -> output-vids/

Roadmap

• More readability, more simplicity. Better code organisation.
• Better documentation.
• Deploy the model.
• Find better data cleaning methods.
  • Upsample 'no-mask' images.
  • Data Augmentation on both sets afterwards. (might increase variance because of inaccurate labels)
• Try and achieve better accuracy on the same dataset. (More epochs, seems to improve)
• Ask senior CV Engineers on the inaccurate label problem. (Segmentation - SAM, etc)
• Write down use cases and try and build niched down cv models.(Reverse blackbox approach)