The Dataset collected from the Kaggle Face Mask Detection Dataset
Here, I was not using the default YOLOv4 model rather using the tiny version of it. The reason behind it the default version takes a huge amount of time to train, where the tiny version takes very little time. However the ability of detection was not compromised where I use the generated anchors on the custom dataset and turn on random initialization to increase the accuracy; finally used 3 YOLO detection layers to detect objects in small, medium, and large scales.
Download the model from [here]