video-grounding

Introduces AnyGroundBench, a domain-adaptation benchmark for spatio-temporal video grounding, evaluating 15 VLMs across five specialized domains and finding current models fail in zero-shot and in-context learning adaptation.