MV-STRIDE — Multi-View Spatial Reasoning QA Dataset

Hierarchical (Level I/II/III) multi-view spatial-reasoning QA data generated by the MV-STRIDE pipeline from two scene sources: Infinigen (synthetic, procedurally generated) and ScanNet++ (real scanned scenes). Every sample asks a multiple-choice spatial question about 2–3 images of the same scene taken from different camera viewpoints.

Layout

• annotations/<source>/<version>_{atomic,stage1,stage3}.json — QA samples; each sample: messages (user: <image><image>Question + Options, assistant: answer), images (relative to this dataset repo, images/<source>/...), category, scene_name, data_source.
• images/<source>/... — the images referenced by the annotations (per-scene dirs mirror the source layout; Infinigen scenes also carry scene_metadata.json, ScanNet++ scenes scene_metadata_new.json).
• configs/ — the exact qa_config_*.json used to generate each split.

Files

source	scenes	file	samples	role
infinigen	439	Infinigen_MultilevelCategories_20260403_sampled_MCA_Multistage_atomic.json	253,400	full set, all levels, single-turn
infinigen	439	Infinigen_MultilevelCategories_20260403_sampled_MCA_Multistage_stage1.json	147,382	SFT split (stage-1 scene pool)
infinigen	439	Infinigen_MultilevelCategories_20260403_sampled_MCA_Multistage_stage3.json	36,631	Level III only, RL split (stage-3 scene pool)
scannetpp	137	ScannetppIphone_MultilevelCategories_20260313_sampled_MCA_Multistage_atomic.json	284,325	full set, all levels, single-turn
scannetpp	137	ScannetppIphone_MultilevelCategories_20260313_sampled_MCA_Multistage_stage1.json	167,889	SFT split (stage-1 scene pool)
scannetpp	137	ScannetppIphone_MultilevelCategories_20260313_sampled_MCA_Multistage_stage3.json	21,608	Level III only, RL split (stage-3 scene pool)

Total: 537,725 single-turn QA samples (Level I perception + Level II cross-view + Level III reasoning incl. MSR).

Load with the datasets library:

from datasets import load_dataset
ds = load_dataset("json", data_files="annotations/*/*_stage1.json")
# images are paths relative to this repo root; join before reading

License & provenance

• QA texts/annotations and Infinigen-derived content: generated by the MV-STRIDE pipeline (cite below); Infinigen is MIT-licensed.
• ScanNet++ images/frames are derived from the ScanNet++ dataset; check and respect its license terms before commercial use / redistribution.

Citation

@inproceedings{xu2026mvstride,
  title     = {MV-STRIDE: Enabling MLLMs to Master Multi-View Spatial Reasoning via Hierarchical Capability Modeling},
  author    = {Xu, Jin and Huang, Xiaojian and Luo, Zhuodong and Zhang, Zhihong and Liu, Xin and Wei, Jiansheng and Wang, Xinzhi and Zhao, Jie and Chen, Xuejin},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}