Robust DeepLearning-Based FaceRecognition Under Unseen Combinations of Real-World Image Distortions

While more recent work evaluates larger OOD situations and realistic degradations, deep face recognition can suffer when capture conditions deviate from training data; blur, noise, and occlusion have negatively impacted performance in face-recognition evaluations.[1-3] Although staged or adaptive training and distortion augmentation are viable approaches, the research addressed here does not prove generalization to systematically held-out combinations of natural distortions.[3, 4] This paper makes the case that robustness evaluation should report task-appropriate, subgroup, and clean-image performance; differentiate distortion types, severities, and combinations; and use naturally acquired data. The main finding is that deployment robustness cannot be fully demonstrated by single-corruption outcomes.