Real time face mask detection with SSD

This paper proposes the Neural Network SSD architecture for the task of face mask detection. We trained this model to detect 3 different categories of people: those wearing a mask, those wearing it incorrectly and those not wearing it. For this proposal, a dataset with 853 images was used with annotations of these 3 classes. For the Neural Network, the weights were initialized from a pretrained model on the image dataset Pascal VOC and fine-tuned to finally achieve a mAP of 70.2% on the validation set. The mAP achieved on the test set was 66.7% which shows that despite the lack of a huge dataset, the architecture SSD is suitable for this task. Additionally, our experiments showed that the MultiBox Loss Function parameters can be modified to improve the performance on small and unbalanced datasets such as the one presented.Current work has focused on the task of classification or detection for a single class. Our work is a novel approach as it performs real-time detection and classification in 3 classes, making it suitable for natural scenarios such as public spaces.