Representing the Unseen: Rethinking Visual Imagination for Spatial Reasoning
Paper • 2610.37437 • Published • 1
Access to ImagiRep requires agreeing to the Matterport3D Terms of Use
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
ImagiRep contains images, videos and geometry rendered from the Matterport3D dataset. Before requesting access, you must have read and agreed to the Matterport3D Terms of Use (https://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf). The data may be used for non-commercial academic research only and may not be redistributed.
Log in or Sign Up to review the conditions and access this dataset content.
ImagiRep
Benchmark data for the paper Representing the Unseen: Rethinking Visual Imagination for Spatial Reasoning (Zengqi Zhao*, Zun Wang*, Shoubin Yu, Mohit Bansal, Yue Zhang; UNC Chapel Hill; * equal contribution).
Project page · Paper (arXiv) · Code and results (GitHub)
ImagiRep has 922 four-option spatial reasoning questions about unobserved viewpoints in Matterport3D scenes:
| setting | questions | groups | scenes | question types |
|---|---|---|---|---|
| Viewpoint Imagination | 622 | 200 trajectories | 73 | Open-space Direction (91), Nearest-obstacle Direction (55), Visibility (104), Nearest Object (104), Farthest Object (98), Object-relative Direction (170) |
| Cross-view Integration | 300 | 143 view groups | 21 | Nearest Object (81), Farthest Object (82), Spatial Ordering (137) |
For every question the imagined (unobserved) state is given as endpoint RGB and six alternative representations of the same state: top-down geometry, spatial text, video, depth, surface normals and panorama. The dataset also holds the estimated representations used in the paper (from HY-World 2.0 for Viewpoint Imagination, and from Stable Virtual Camera with MoGe-2 depth for Cross-view Integration).
Contents
| path | content |
|---|---|
benchmark/benchmark922/questions.jsonl |
922 rows: question, options, gold answer, system prompt, and the evidence of every input condition |
benchmark/benchmark922/tasks.jsonl |
7,376 model-ready requests (922 questions x 8 conditions) as ordered parts {"text"}, {"image"}, {"video"} |
benchmark/benchmark922/images/, videos/ |
all images and videos referenced by the two files above; SHA256SUMS |
benchmark/uid_lists_and_index/ |
uid lists and question_index.tsv (category, gold, trajectory or view group, scan, look-back flag) |
benchmark/camera_poses/ |
camera poses and Matterport3D scan / viewpoint ids for every question |
benchmark/randomization/ |
how each question was randomized from its annotation (option order; image numbering for Cross-view Integration) |
03_estimated_viewpoint_gpt61sol/ |
estimated representations for Viewpoint Imagination: estimated_assets/ and the exact request images gpt_runs/images/ |
04_estimated_crossview_gpt61sol/ |
estimated RGB and depth for Cross-view Integration: gpt_runs/inputs/ and gpt_runs/images/ |
The folders follow the layout of the GitHub repository: clone it and download this dataset into its root, and the
scripts there find every file (hf download Imagirep/ImagiRep --repo-type dataset --local-dir .).
The request files and per-question model outputs that use the estimated images are in the GitHub repository.
Input conditions: observed (observed view(s)), oracle (+ endpoint RGB; for look-back questions a view from the
endpoint facing the start) and oracle_<m> (+ one representation m: topdown, text, video, depth,
surface_normals, panorama).
import json
qs = [json.loads(l) for l in open("benchmark/benchmark922/questions.jsonl")]
q = qs[0]
print(q["uid"], q["category"], q["gold"])
print(q["question_prompt"])
print(q["conditions"]["oracle_panorama"]) # text parts and file names in images/ and videos/
License and terms
See LICENSE.md. Images, videos and geometry are derived from Matterport3D and remain subject to the Matterport3D
Terms of Use: non-commercial academic research only, no redistribution. Questions, labels and other annotations are
released under CC BY-NC 4.0.
Citation
@article{zhao2026imagirep,
title = {Representing the Unseen: Rethinking Visual Imagination for Spatial Reasoning},
author = {Zhao, Zengqi and Wang, Zun and Yu, Shoubin and Bansal, Mohit and Zhang, Yue},
journal = {arXiv preprint arXiv:2610.37437},
year = {2026}
}
- Downloads last month
- -