Overview
Along with practicing social distancing, wearing a face mask is an effective way to keep our community safe during the COVID-19 pandemic. Researchers have shown that wearing a mask can reduce infection risk to the wearer by 65%. The importance of face mask is more important than ever before, with more states mandating face masks. However, face masks are most effective when everyone wears one, and it is impractical to ensure everyone is wearing a mask by sticking to traditional methods and older technologies.
At Lanthorn.ai, we leverage AI to help businesses get back to work safely during the COVID-19 pandemic. Our Face Mask Detector application can detect face masks even in low-resolution CCTV videos, and it works in real-time. This application is compatible with the existing cameras installed at hospitals, workplaces, schools, and elsewhere. The best part about our application is that no personal data is stored anywhere; the computations take place locally, at the edge of the network, and users’ privacy is completely preserved. Our model was purposefully developed to run on light weight edge devices such as the Jetson Nano and Google Coral.
Accuracy & Speed Assessment
To the best of our knowledge, our face mask classifier outperforms all currently available face mask classifiers that can work on CCTV video streams. We are more accurate than any other published models in detecting if a person is wearing a mask. We are also one of the fastest! We can classify two thousand (2K) faces per second on TPU. Due to the speed of our model, our classifier remains rapid even when run simultaneously with other models.
Model Training for Accuracy
To train our face detection model, we collected data, annotated the data, and trained our model. To gather the dataset, we combined datasets from Wider Face, CelebA, and SM-Synthetic results in 23K images. We then dedicated an immense amount of time and effort to label this dataset for “mask/no mask” classification. Subsequently, a dataset of 50K “mask/ no mask” images created from synthesizing 52 types of face masks on faces (available for download here).
To train a face mask detector, our R&D team conducted several experiments to obtain the best model architecture with the highest accuracy. Our major challenge was detecting face masks in videos captured by CCTV cameras, where faces only occupy a small portion of the video frame and are low-resolution and blurry.
We mitigated this challenge by training our face mask classifier on FaceMask100K, a large-scale, diverse “mask/ no mask” dataset, collected by Neuralet (our R&D arm). We also used pose estimation algorithms to detect faces in CCTV video streams more accurately. You can read our face mask article to learn more about our Face Mask Detector application. The video below is a great showcase of Lanthorn.ai’s speed and accuracy.
Interested in Lanthorn.ai?
If you are interested in testing our mask detection model in your environment contact sales@lanthorn.ai
