Extending Earth Mover's Distance to Occluded Face Verification

Facial recognition is one of the most used biometric clues to identify and verify individuals regarding access to secure facilities or devices, for law enforcement purposes, or to locate missing persons. This is due to its non-intrusive nature and high distinguishable power for identity authentication. Deep learning systems have been shown to be the go-to choice to extract distinguishable features from faces. However, a common real world scenario challenge occurs when occlusions (eg. sunglasses, masked, scarf) in the input to such systems decrease their overall performance. To address this issue, the scientific community proposed approaches and competitions, resulting in different solutions to this problem. We focus on extending the Earth Mover's Distance (EMD) to the occluded face recognition problem by evaluating its potential with state-of-the-art backbones on verification tasks by using the OCFR-2022 benchmark. We confirm the applicability of fusing the cosine similarity and EMD distance scores to enhance conservative decision-making process. We draw this conclusion by considering lowering the False Match Rate operation points on verification set.