Abstract

SAIL tackles weakly-supervised dense video captioning through similarity-aware guidance and inter-caption augmentation. The framework reduces reliance on dense temporal annotations while maintaining strong captioning quality.