The Dataset used in this project is available in Kaggle and can be downloaded from here.
It consists of images and corresponding labels for three different classes as follows:
- With Mask ["WM"]
- Without Mask ["WoM"]
- Mask wore incorrect ["MWI"]
The given labels are in Pascal VOC format (XML) they should be converted into YOLO format (Text) for training. After that data is split into 80:20 ratio.
The training for this notebook is done on Google Colaboratory.
The model is trained using the given dataset for ~120 epochs. Other parameters are
Image size : 640 x 640
Batch size : 32
Class/Metric | Precision | Recall | mAP50 | mAP50-95 | F1 Score | Accuracy |
---|---|---|---|---|---|---|
All | 0.816 | 0.764 | 0.831 | 0.560 | 0.7891 | 92.45% |
WM | 0.944 | 0.908 | 0.956 | 0.655 | 0.9256 | 88.06% |
WOM | 0.787 | 0.752 | 0.790 | 0.475 | 0.691 | 91.39% |
MWI | 0.717 | 0.633 | 0.747 | 0.551 | 0.6723 | 97.84% |