Event-STVG: Spatio-Temporal Video Grounding for Long-Form Event Reasoning

一句话要点

清华的 Event-STVG 把视频时空定位从“找描述的目标”推进到长视频事件推理，涵盖重复事件、角色推断和多目标定位，基准含 4,016 个事件标注，最强模型也很吃力。

Tsinghua's Event-STVG moves spatio-temporal video grounding toward long-form event reasoning — repeated events, role inference, multi-target grounding — with a 4,016-event benchmark where top MLLMs still struggle.