Tag
Introduces PEEK, an efficient dynamic frame sampling method that distills caption-conditioned frame relevance rankings from a teacher model into a lightweight temporal model, outperforming state-of-the-art methods in video captioning while maintaining computational efficiency.