A Robust System to Detect and Explain Public Mask Wearing Behavior

Abstract

COVID-19 is a global health crisis during which mask-wearing has emerged as an effective tool to combat the spread of disease. During this time, non-technical users like health officials and school administrators need tools to know how widely people are wearing masks in public. We present a robust and efficient Mask Adherence Estimation Tool (MAET) based on the pre-trained YOLOv5 object detection model and combine it with explanation methods to help the user understand the mask adherence at an individual and aggregate level. We include two novel explanation methods to compute a high-fidelity importance map based on two black-box explanation methods. For our work, we experimented with one-stage and two-stage object detector architectures. Experiment results show that MAET achieves state-of-the-art results on a public face mask dataset, with improved performance by 2.3 % precision and 0.4 % recall in face detection and 2.0 % precision and 1.7 % recall in mask detection. We used three different evaluation metrics for explanation and find that no method dominates all metrics; therefore, we support multiple explanation methods.