Tag
The paper introduces HN-CLIP, a method that uses the text encoder's text-text geometry to create adaptive similarity margins for dense-caption retrieval, addressing saturation issues in contrastive learning and improving performance over existing methods.