RefSpatial

From the catalogue. Listed by one researcher from one source page.

• id: refspatial
• kind: language and reasoning annotation
• robot or device: OpenImages 2D images, CA-1M 3D videos, Blender simulated data (class: sensor only)
• how collected: generated and filtered annotation
• size as stated: 20M QA pairs (paper); 466k images, 100k video frames, 3k annotated assets; ~357GB raw
• hours: not stated | hours only claimed: none | episodes: not stated
• year: 2025
• organisation: JingkunAn (HF account)
• licence: apache-2.0 (class: open)
• access: open (class: open)
• link (dataset): https://huggingface.co/datasets/JingkunAn/RefSpatial
• Hugging Face repository: JingkunAn/RefSpatial, 6,014 downloads in 30 days, https://huggingface.co/datasets/JingkunAn/RefSpatial
• note: Training set for RoboRefer spatial referring; 31 spatial relations.

Confirm the licence at the link before relying on it.