Mind-the-Gap! Unsupervised Domain Adaptation for Text-Video Retrieval

When can we expect a text-video retrieval system to work effectively on datasets that differ from its training domain? In this work, investigate question through the lens of unsupervised domain adaptation in which objective is match natural language queries and video content presence shift at query-time. Such systems have significant practical applications since they are capable generalising new data sources without requiring corresponding text annotations. We make following contributions: (1) propose UDAVR (Unsupervised Domain Adaptation for Video Retrieval) benchmark employ it study performance shift. (2) Concept-Aware-Pseudo-Query (CAPQ), method learning discriminative transferable features bridge these cross-domain discrepancies enable effective target using source supervision. (3) show CAPQ outperforms alternative strategies UDAVR.