Revisiting Uncertainty: On Evidential Learning for Partially Relevant Video Retrieval
Researchers have developed Holmes, a new framework for partially relevant video retrieval that explicitly models uncertainty. This hierarchical evidential learning approach aggregates evidence across different granularities to handle the ambiguity between brief text queries and extensive video content. Holmes uses Dirichlet distributions to interpret similarity scores and employs optimal transport for query-clip alignment to improve retrieval accuracy, outperforming existing methods. AI
IMPACT Introduces a novel method for handling uncertainty in video retrieval, potentially improving search accuracy for complex, partially described content.
