这是一个从 DROID / RoboInter 操作数据中筛选并人工复核的真实机器人多视角分割数据集。给定一个
视角中的目标物体图像与 mask,任务是在同一时刻的另一个相机视角中分割同一物体。数据集保留
ego(wrist)与两个第三人称视角,并记录每个最终 mask 的来源及人工修订过程。
This dataset evaluates cross-view object correspondence in real robot manipulation:
given an object mask in a source camera, segment the same physical object in a
synchronised target camera. The source material is the DROID subset distributed
through RoboInter; target identities and… See the full description on the dataset page:
https://huggingface.co/datasets/Travor278/droid-crossview-seg.