Self-supervised Rare Visual Token Enhancement
Adrien Chan-Hon-Tong · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
This paper focuses on rare visual token enhancement: given an image the model is trained to select few visual tokens with the property of being largely distant from all the others.This idea, which is structurally more frugal than classical self-supervised methods, could be an interesting pretext task for both fast classification based on few tokens, or, interesting point selection.