Tag
The paper presents EgoCaptioner, a coarse-to-fine egocentric video captioning framework that combines video-level overviews with timestamped, perception-adaptive step narration, trained with GRPO using factual and temporal rewards, achieving leading results on five egocentric benchmarks and surpassing proprietary baselines like Gemini 3.1 Pro on PhysCapBench.
Temporal Context Routing improves script-aligned timing in joint audio-video generation by mapping script timing onto shared video-audio temporal axes, reducing errors in shot boundaries and dialogue.
EgoCS-400K is a large-scale egocentric Counter-Strike dataset with over 400,000 first-person videos and 10,000 hours of gameplay, providing temporally aligned video-action-language trajectories for world model research.