A face detection ensemble to monitor the adoption of face masks inside the public transportation during the COVID-19 pandemic

The designing of ensembles is widely adopted when single machine learning methods fail to obtain satisfactory performances by analyzing complex data characterized by being imbalanced, high-dimensional, and noisy. Such a failure is a well-known statistical challenge when the learning algorithm searches for a model in a large space of hypotheses and the data do not significantly represent the problem, thus not inducing it from a space of admissible functions towards the best global model. We have addressed this issue in a real-world application, whose main objective was to identify whether users were wearing masks inside public transportation during the COVID-19 pandemic. Several studies have already pointed that face masks are an important and efficient non-pharmacological strategy to reduce the virus spread. In this sense, we designed an approach using Convolutional Neural Networks (CNN) to track the adoption of masks in different transportation lines, regions, days, and time. Aiming at reaching this goal, we propose an ensemble of face detectors and a CNN architecture, called MaskNet, to analyze all public-transport passengers and provide valuable information to policymakers, which are able to dedicate efforts to more effective advertisements and awareness work. In practice, our approach is running in a real scenario in Salvador (Brazil).