Internal research and participation in international research projects involve the following subjects:
- deep learning techniques
- identification from photographs or video data using 3D facial images
- face recognition with low-resolution images and video material
- pose variation and partial occlusion in facial images
- tracking people in crowds, even when the face is not visible
In light of recent global developments, researchers at Cognitec have investigated the impact of face masks on face recognition accuracy, especially for unsupervised scenarios, where many faces are being matched against large databases. As expected, our current technology matches masked faces, but tests showed an overall decrease of face detection rates and an increase of false match rates, in comparison to unmasked faces.
Some face recognition companies are publishing doubtful claims about their technologies reaching the same matching accuracy for faces covered by masks. A report published by the National Institutes of Science and Technology (NIST) in July shows the performance of 89 face recognition algorithms when matching faces with digitally applied face masks with unmasked faces of the same person, with error rates between 5% and 50% . Cognitec did not participate in this test.
Since masks occlude a large area of the face, removing substantial information needed for comparison tasks, matching accuracy will evidently decrease. Therefore, unmasked faces will remain the gold standard for automated face recognition technologies. High-security applications that need to establish a person's identity, for example, border control or passport issuance, will most likely require the removal of masks. But as long as wearing masks is required or common in everyday life, recognizing masked faces is becoming an expected feature for systems that unlock phones, track and count people, measure demographics, or identify persons of interest.
Cognitec is responding to such market demands as we continuously improve our algorithms. Our research and development team constantly works on increasing the accuracy for partially occluded faces. In addition, we are using training data with masked faces, and current lab results already show significant advances for detecting and matching faces with masks. Stand by for product releases!