QianYangMILA/vdrop-crossview-8k

Cross-View Visual Thinking — 8K Training Set

Training data for “How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning”.

This repository contains the training set only (single train split). Evaluation benchmarks used in the paper are third-party and are not distributed here.

Given two egocentric views of the same indoor scene with partially overlapping fields of view, each sample asks a multiple-choice cross-view spatial question. Uniquely, every scene ships with ground-truth renders of the candidate thinking-images — panorama, corner-view (“top-down” in the paper), point-matching, and a BEV ortho top-down — enabling supervised visual chain-of-thought training and controlled comparison across intermediate representations.

All scenes are procedurally generated with Infinigen Indoors, so answers and thinking-images are unambiguous by construction.

Contents

QA pairs	7,728 (intersection of all four thinking-image variants)
Unique scenes	1,255
Question types	anchor (716) · counting (1,141) · relative-distance (2,961) · relative-direction (1,458) · perspective-taking (1,452)
Images per scene	6 — cam0, cam1, panorama, corner_view, pm_bridge, topdown_blender

Config qa — one row per question

id (<question_type>_<NNNNNN>), source (constant infinigen), scene_id, question_type, question (options inline), answer, plus repo-relative paths: cam0, cam1, bridge_panorama, bridge_corner_view, bridge_pm, bridge_topdown_blender.

Config scenes — one row per scene, images embedded

source, scene_id + the six images as Image features (viewer-friendly; join to qa on scene_id).

bagel_sft/ — ready-to-train JSONLs (ShareGPT schema)

The exact files used to fine-tune BAGEL in the paper: no_thinking.jsonl plus visual_only_{panorama,corner_view,pm,topdown}.jsonl (fields: conversations, image, visual_thinking, id; image paths are images/<scene_id>/<name>.png, relative to the repo root). Pair with the training code in the paper’s code release. Run materialize_images.py first (below) so those paths resolve.

Images — embedded in scenes, no loose PNG tree

The images are shipped inside the scenes config (embedded bytes), not as a loose images/ folder — uploading thousands of tiny PNGs trips the Hub's request-rate limit. To materialize the exact tree the qa / bagel_sft paths expect (images/<scene_id>/{cam0,cam1,panorama,corner_view,pm_bridge,topdown_blender}.png), run the included materialize_images.py once — it decodes scenes to disk (byte-identical, resumable):

pip install "datasets>=2.16"
python materialize_images.py --repo_id QianYangMILA/vdrop-crossview-8k          # writes ./images/

Usage

from datasets import load_dataset

# text questions (image columns are relative paths)
qa = load_dataset("QianYangMILA/vdrop-crossview-8k", "qa", split="train")

# scenes with the 6 images already decoded (no separate download needed)
scenes = load_dataset("QianYangMILA/vdrop-crossview-8k", "scenes", split="train")
scenes[0]["panorama"]            # -> PIL.Image

# to get a loose images/<scene_id>/*.png tree (for the SFT JSONL paths):
#   python materialize_images.py --repo_id QianYangMILA/vdrop-crossview-8k

Notes

• The paper reports 7,921 training traces; this release keeps the 7,728 ids for which all four thinking-image renders exist (intersection-first), so every variant covers identical questions.
• Rendered with Blender from procedurally generated Infinigen Indoors scenes.

Citation

@article{yang2026and,
  title={How and What to Imagine? Visual Thinking in Unified Multimodal Models for Cross-View Spatial Reasoning},
  author={Yang, Qian and Sikarwar, Ankur and Le, Huy and Zhang, Le and Shi, Zhuan and Taslakian, Perouz and Agrawal, Aishwarya},
  journal={arXiv preprint arXiv:2605.27310},
  year={2026}
}

Please also cite Infinigen (scene generation).