Pokerme/view2space-train

VIEW2SPACE Training

This repository contains the public VIEW2SPACE training release for multi-view vision-language spatial reasoning. [image] Associated paper:

• —VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations - ECCV 2026 🚀
• —arXiv: 2603.16506
• —Project Page: Project Page

The training release contains:

• —604,779 training examples
• —22,205 public images
• —three public question families: count, detect, and mcq

Unlike the evaluation release, this training package keeps supporting.chain_of_thought so it can be used for supervision. [image]

Related VIEW2SPACE Releases

• —Testing release: Pokerme/view2space-v1
• —4B model checkpoint: Pokerme/view2space_4b

For the official project, updates, and citation information, see:

VIEW2SPACE GitHub repository

Directory Structure

view2space-train-release/
  README.md
  overall.jsonl
  images/
    img_000001.png
    img_000002.png
    ...

Each line in overall.jsonl is one training example.

Data Format

Each JSONL record contains the following public fields:

• —q_idx: anonymized question id such as count_000001
• —q_type: public question family, one of count, detect, or mcq
• —question: question text
• —options: multiple-choice options for MCQ questions; empty for non-MCQ questions
• —question_prompt: extra prompt text when applicable
• —answer: ground-truth answer
• —image_paths: relative paths to the images used by the question
• —supporting.draw_boxes: optional input boxes associated with one or more images
• —supporting.chain_of_thought: released reasoning trace for training

Notes:

• —image_paths are relative to the dataset root.
• —Example image path: images/img_000123.png
• —original internal scene names, image names, and fine-grained question types are anonymized in the public release

Example Record

{
  "q_idx": "mcq_000001",
  "q_type": "mcq",
  "question": "Based on these four images (image 1, 2, 3, and 4) showing the ball from different viewpoints (front, left, back, and right), with each camera aligned with walls and partially capturing the surroundings: \nFrom the viewpoint presented in image 4, what is to the left of ball?",
  "options": {
    "A": "tv",
    "B": "fountain",
    "C": "sofa",
    "D": "table"
  },
  "question_prompt": "",
  "answer": "B",
  "image_paths": [
    "images/img_000001.png",
    "images/img_000002.png",
    "images/img_000003.png",
    "images/img_000004.png"
  ],
  "supporting": {
    "draw_boxes": null,
    "chain_of_thought": "<think> ... </think>\n <answer> B </answer>"
  }
}

Basic Usage

import json
from pathlib import Path

dataset_root = Path("/path/to/view2space-train-release")
jsonl_path = dataset_root / "overall.jsonl"

with jsonl_path.open("r", encoding="utf-8") as f:
    first = json.loads(next(f))

image_files = [dataset_root / rel_path for rel_path in first["image_paths"]]
reasoning = first["supporting"]["chain_of_thought"]

print(first["q_idx"])
print(first["q_type"])
print(image_files)
print(reasoning[:200])

Citation

License and Citation

This dataset is licensed under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Attribution is required under the license. For academic and research use, citation in the resulting paper, report, model card, dataset card, or public documentation is expected. Please cite this dataset whenever it is used for training, evaluation, benchmarking, data analysis, or as part of a larger dataset mixture.

@article{ke2026view2space,
  title={VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations},
  author={Ke, Fucai and Cai, Zhixi and Li, Boying and Chen, Long and Lin, Beibei and Wang, Weiqing and Haghighi, Pari Delir and Haffari, Gholamreza and Rezatofighi, Hamid},
  journal={arXiv preprint arXiv:2603.16506},
  year={2026}
}