In short, this work studies how to make VideoLLMs more camera-aware by benchmarking camera motion understanding and injecting geometry-derived motion cues at inference time.
Mar 13, 2026
