EgoExo-Con: Exploring View-Invariant Video Temporal Understanding

The article introduces EgoExo-Con, a benchmark designed to assess the temporal understanding of Video-LLMs using synchronized egocentric and exocentric video pairs. It highlights two key tasks: Temporal Verification and Temporal Grounding, revealing that existing models struggle with consistency across viewpoints, often performing worse than single-view scenarios. The proposed View-GRPO framework employs reinforcement learning to enhance view-specific temporal reasoning and improve cross-view consistency, addressing critical limitations in current Video-LLMs.