VEDJE: Video-Efficient Discriminative Joint Encoder for Scalable Video-Text Retrieval
VEDJE addresses needing a reusable video cache that stores evidence for various objects and actions while remaining small enough for large-scale storage. It achieves this by compressing features within sampled frames offline, creating a compact representation reused at query time alongside text and a score. The system uses feature-change supervision to improve retrieval under tight storage budgets.