view2space-train
收藏资源简介:
VIEW2SPACE训练数据集是一个专门用于多视角视觉语言空间推理的大规模公开数据集,包含604,779个训练示例和22,205张公开图像,旨在支持多视角视觉推理任务的研究。数据集涵盖三种公开问题类型:计数(count)、检测(detect)和多项选择题(mcq)。与评估版本不同,训练版本保留了支持推理链(chain_of_thought),可用于监督学习。每个训练示例以JSONL格式存储,包含匿名化的问题ID、问题类型、问题文本、多项选择选项(仅限MCQ问题)、正确答案、图像路径(相对路径)以及支持信息(包括可选的绘制框和推理链)。图像路径指向数据集根目录下的images文件夹。该数据集适用于视觉问答、空间推理、多模态理解等任务的研究和模型训练。
The VIEW2SPACE training dataset is a large-scale public dataset specifically designed for multi-view visual-language spatial reasoning. It contains 604,779 training instances and 22,205 public images, and aims to support research on multi-view visual reasoning tasks. The dataset covers three public question types: count, detect, and multiple choice question (MCQ). Unlike the evaluation split, the training split retains chain-of-thought support, which can be used for supervised learning. Each training instance is stored in JSONL format, including anonymized question ID, question type, question text, multiple choice options (only for MCQ questions), correct answer, image path (relative path), and supporting information (including optional bounding boxes and chain-of-thought). The image path points to the images folder under the root directory of the dataset. This dataset is applicable to research and model training for tasks such as visual question answering, spatial reasoning, and multimodal understanding.
数据集概述:VIEW2SPACE Training
VIEW2SPACE Training 是一个用于多视角视觉-语言空间推理的训练数据集,旨在通过稀疏观察研究多视角视觉推理。
核心信息
- 数据集名称:VIEW2SPACE Training
- 任务类别:视觉问答 (Visual Question Answering)
- 关联论文:VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations (ECCV 2026)
- arXiv:2603.16506
- 项目主页:Project Page
数据规模与构成
- 训练样本:604,779 个
- 公开图片:22,205 张
- 问题类型:包含三种公开问题族:
count(计数)detect(检测)mcq(多选题)
注意:训练包保留了 supporting.chain_of_thought 字段,可用于监督训练。
目录结构
view2space-train-release/ README.md overall.jsonl images/ img_000001.png img_000002.png ...
数据格式
每条训练样本以 JSONL 格式存储,每条记录包含以下字段:
| 字段 | 说明 |
|---|---|
q_idx |
匿名化的问题ID,例如 count_000001 |
q_type |
问题类型:count、detect 或 mcq |
question |
问题文本 |
options |
多选题选项(仅 MCQ 问题有效,非 MCQ 问题为空) |
question_prompt |
额外提示文本(如有) |
answer |
标准答案 |
image_paths |
相关图片的相对路径列表 |
supporting.draw_boxes |
可选输入框(与部分图片关联) |
supporting.chain_of_thought |
用于训练的推理过程记录 |
说明:
image_paths相对于数据集根目录,例如images/img_000123.png- 原始场景名称、图片名称及细粒度问题类型已在公开版本中匿名化
示例记录
json { "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: From 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> <answer> B </answer>" } }
许可与引用
- 许可证:Creative Commons Attribution 4.0 International License (CC BY 4.0)
- 预期用途:学术与研究用途,用于训练、评估、基准测试、数据分析或作为更大数据集的一部分时,需引用本数据集。
- 引用格式:
bibtex @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} }
相关发布
- 测试集:Pokerme/view2space-v1
- 模型检查点:Pokerme/view2space_4b
- GitHub 仓库:VIEW2SPACE GitHub repository




