Face presentation attack detection—the task of telling a real face from a photo, video, or mask—has been studied mainly under fixed camera or handheld acquisition. A new arXiv paper argues that smart glasses create a fundamentally different setting, one that existing methods may not handle well.
The authors point to specific difficulties introduced by smart glass capture: ego motion as the wearer moves, constantly changing viewpoints, cluttered backgrounds, and varying distance to the subject. These conditions break the assumptions of controlled or semi-controlled capture that most prior work relies on.
To tackle this, the paper proposes a decoupled representation learning approach, which aims to separate the factors that distinguish genuine faces from presentation attacks. The abstract does not include experimental results, so the effectiveness of the method is not yet reported in the source.