VidSTG Unified Review
splittrain
shown10 / 10
SAM3.1 masks模型结果 · 只读
trainVLM 0.5 FPSread-only prompt-aligned

Video evidence

模型发送给 VLM 的 clean frames;点击缩略图跳转。
当前视频证据帧

Mask identity sheet

对象身份参考图

Caption · objects & actions

Actions

action text · roles · mapped time

Entities & grounding

all mentions: click row or object card to select marks