VISTA (Video Grounding) - Human-Human: leaderboard

VISTA (2026) spatio-temporal video grounding on queries describing interactions between people. Mean spatio-temporal IoU in percent; higher is better.

Metric: Mean spatio-temporal IoU (m_vIoU, %): per-frame box IoU summed over the frames where predicted and true tubes overlap in time, divided by their temporal union, on human-human interactions, over the 11,814 video-query pairs of VISTA (HCSTVG-v1 and v2, VidVRD, VidSTG, MeViS and RVOS videos re-annotated with an interaction taxonomy), zero-shot on sub-sampled frames; the model must localize the queried subject with a box in every frame; higher is better. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 10 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B66.2#917
2InternVL2.5-8B49.2#2370
3Qwen-VL-Chat48.2#7860
4VISTA Sphinx-v2 (checkpoint unspecified)47
5MiniGPT-v246
6VISTA MiMo-VL-7B (checkpoint unspecified)45.1
7VISTA CogVLM Grounding (checkpoint unspecified)44.7
8Shikra-7B34
9LLaVA-Grounding-7B28.4
10Ferret-7B26.4

Interactive version: theaggregate.ai/benchmark?slug=vista-video-grounding-human-human · How It Works · Data refreshed daily, snapshot 2026-11-21.