NearID: Identity Representation Learning via Near-Identity Distractors

Aleksandar Cvejic, Rameen Abdal, Abdelrahman Eldesokey, Bernard Ghanem, Peter Wonka

Abstract

Trains identity-aware visual representations by contrasting each identity against visually similar but distinct "near-identity" instances, separating true identity signal from background context and improving both embeddings and identity-based evaluation metrics.