AnyGroundBench: A Specialized-Domain Benchmark for Video Grounding in Vision-Language Models

AnyGroundBench reframes spatio-temporal video grounding evaluation as a domain-adaptation problem rather than a static zero-shot test on general-purpose datasets. It covers five specialized domains: animals, industry, sports, surgery, and public security. The benchmark combines newly collected videos, including expert-annotated mouse behavior footage, with existing datasets and dense spatio-temporal annotations. It also provides training splits for measuring adaptation. Evaluations of 15 state-of-the-art VLMs examine both zero-shot generalization and in-context learning under practical compute constraints. The reported results indicate that current models struggle with both forms of adaptation in specialized settings.