VEDJE: Video-Efficient Discriminative Joint Encoder for Scalable Video-Text Retrieval
VEDJE keeps video-search accuracy high while cutting the reusable per-video cache down to compact evidence.
The paper describes a joint video-text retrieval encoder that compresses sampled-frame features but keeps them separately cached for later queries. An auxiliary feature-change prediction signal improves retrieval from that compressed cache without adding query-time cost. Across MSR-VTT, MSVD, DiDeMo, and ActivityNet, it reports better R@1 than matched first-stage retrievers in both retrieval directions. In one VideoPrism setup, a fourfold cache reduction to 12 KiB stays within 0.2 text-to-video recall points.
HF Daily Papers' note
The paper describes a joint video-text retrieval encoder that compresses sampled-frame features but keeps them separately cached for later queries. An auxiliary feature-change prediction signal improves retrieval from that compressed cache without adding query-time cost. Across MSR-VTT, MSVD, DiDeMo, and ActivityNet, it reports better R@1 than matched first-stage retrievers in both retrieval directions. In one VideoPrism setup, a fourfold cache reduction to 12 KiB stays within 0.2 text-to-video recall points.
HF Daily Papers' note
score 4