Think Visually Across Space and Time: Grounded Reasoning for Video Segmentation

One-line takeaway

For video reasoning segmentation — finding an implicitly described target and predicting its mask trajectory — this paper proposes Temporal Chain of Video Primitives (TCoVP), a spatiotemporally grounded video chain-of-thought that orders query-relevant observations chronologically, tying each step to visual evidence that text-only CoT leaves implicit.

这篇论文做视频推理分割：要根据隐含描述找出目标并输出逐帧掩码。作者提出时序视频原语链（TCoVP），把与问题相关的观察按时间排成推理链，每一步都绑定到具体帧里的目标，弥补纯文本思维链看不到视觉证据的问题。