Text-to-video retrieval models struggle with complex queries

A new paper analyzes the performance plateau in text-to-video retrieval systems, evaluating 14 state-of-the-art methods across three datasets. The research found that simpler, clearer captions describing single actions or attributes yield higher retrieval recall. Complex events and multi-step activities remain challenging for current models, with attention-driven architectures showing an advantage for temporally dependent queries. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Identifies key dataset factors and query complexities that hinder text-to-video retrieval, guiding future model development.

RANK_REASON This is a research paper published on arXiv analyzing existing text-to-video retrieval methods. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CV →

• arXiv
• Björn Þór Jónsson
• Hugging Face

• paper
• other