SpatialUncertain

A controlled 3D benchmark evaluating whether Vision-Language Models (VLMs) know when not to answer spatial reasoning questions under occlusion and perspective ambiguity.

Companion to the paper Seeing Isn't Knowing: Do VLMs Know When Not to Answer Spatial Questions (and Why)?

Structure

SpatialUncertain/
├── questions/
│   ├── occlusion_benchmark.json         # ~6,600 questions
│   ├── perspective_benchmark.json       # ~3,700 questions
│   └── view_selection_benchmark.json    # 670 questions
└── scenes/                              # downloaded from the tar.gz files below
    ├── clean_scenes/                    # 555 base scenes
    ├── occlusion_scenes/                # 272 occluded variants
    ├── perspective_scenes/              # 505 perspective-ambiguous variants
    └── view_selection_scenes/           # 505 multi-view variants

Each scene folder contains:

• multiview/view_XXX.png — 8 rendered viewpoints
• camera_poses.json — per-view camera parameters
• object_attributes.json — object metadata
• structure_proxy.json, target_subgraph.json — scene-graph annotations
• walkthrough.mp4 — a short walk-through render

Files

File	Size	Contents
questions/*.json	17 MB	All questions, ground-truth answers, condition tags
clean_scenes.tar.gz	…	Base scenes (no challenge)
occlusion_scenes.tar.gz	…	Occlusion variants
perspective_scenes.tar.gz	…	Perspective-ambiguity variants
view_selection_scenes.tar.gz	…	View-selection variants

Quick start

from huggingface_hub import snapshot_download
# Questions only (small, ~17 MB)
snapshot_download("zhangyuejoslin/SpatialUncertain",
                  repo_type="dataset",
                  allow_patterns="questions/*")

# Full dataset (large)
snapshot_download("zhangyuejoslin/SpatialUncertain", repo_type="dataset")

After downloading, extract the tars:

for f in *_scenes.tar.gz; do tar -xzf "$f"; done

License

Released under CC-BY-4.0. Please cite the paper if you use this benchmark.

Citation

@article{zhang2026spatialuncertain,
  title  = {Seeing Isn't Knowing: Do VLMs Know When Not to Answer Spatial Questions (and Why)?},
  author = {Zhang, Yue and Wang, Zun and Lin, Han and Bitton, Yonatan and Szpektor, Idan and Bansal, Mohit},
  journal= {arXiv preprint arXiv:XXXX.XXXXX},
  year   = {2026}
}