Vintra Unveils New COVID-19 Feature: Face Covering Detector

Reopening Safely

The recent increase in coronavirus infections in the U.S. is alarming. It underscores the need for organizations and businesses to deploy technologies that rebuild trust in the physical spaces that they operate. Vintra’s new face covering detector is a great solution to enable organizations to maintain safety protocols and build trust with staff and guests. The new feature can be used to search for and do real-time alerting on people not wearing a face covering. The face covering detector will start to ship in Q3, and be available in Vintra’s FulcrumAI Investigator (cloud) and FulcrumAI Real Time on-premise solutions.

Today we know that wearing a face covering and social distancing are some behavioral actions that can help to mitigate the spread of the coronavirus. For the foreseeable future this is going to be a standard requirement in places where we work, play, and learn. Security professionals leveraging FulcrumAI Real Time can increase situational awareness across their entire environment and proactively manage new use cases associated with the coronavirus.

Face Covering Detection Technology

As with most machine learning models for object detection, the process of creating a new algorithm begins with sourcing training data. Dr. Amato and the team deployed a transfer learning technique that utilizes both real and synthetic data to train new models. We leveraged our Analytics Foundry platform to create training data and quickly spin up two new models. The first approach detects a person with a face covering based on an overall body detection. The second approach detects a face covering based on the face detection. We then combined both approaches to increase the accuracy. At the end of the day, if we can detect a face we’ll detect a face covering, too.

Example faces of real and synthetic persons. Faces may have varying degrees of occlusion and types of masks or face coverings, thus making their detection challenging.

Performance for Detecting Face Coverings

Our face coverings detector achieves an average precision of 95% on our validation set. Our goal was to achieve 90% accuracy for face covering detection before implementing the feature. Once we repeatedly achieved a high performance score in the synthetic environment we used transfer learning to move the model to production and began testing in the real world. Testing and tuning of the data models will continue to adapt to changes in the various types of face coverings, CDC guidelines, and mandates at the state and local levels.