SS-Faster-RCNN: A Domain Adaptation-based Method to Detect Whether People Wear Masks Correctly

Since the outbreak of coronavirus (COVID-19) in 2019, wearing masks has been widely considered an effective method of reducing the risk of infection among people. However, incorrectly wearing a mask can significantly increase the risk of spreading the virus. To enable machines to automatically detect whether people are wearing masks correctly, we propose scenario-specific Faster-RCNN (SS-Faster-RCNN), a domain adaptation-based method for masked face detection. The frame-work is based on the Faster-RCNN and consists of two parts. The first part detects mask-wearing zones, and the last part aims to validate real mask faces in the candidate regions. The experimental results demonstrate that a trained feature extractor on a large mask face dataset can effectively enhance the model's performance on smaller mask face datasets for scenario-specific and non-scenario-specific cases. In addition, we also present videos for mask detection that consist of 858 seconds of video with 30 frames per second. The dataset mainly consists of images and videos from streets and subway stations. Overall, our method shows superior performance compared to others across different datasets. Codes, data, and evaluations are available at https://github.com/boranyang-ML/SSMFVD.