view2space-v1
收藏资源简介：
VIEW2SPACE v1 是一个用于空间推理的多视图视觉语言评估数据集。该数据集与题为《VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations》的 ECCV 2026 论文相关联。数据集公开发布版本包含三个子集：`count`（计数）、`detect`（检测）和 `mcq`（多项选择题）。数据以 JSONL 格式组织，每个文件（如 `overall.jsonl`）的每一行代表一个问题。每个数据记录包含以下字段：问题ID (`q_idx`)、问题类型 (`q_type`)、问题文本 (`question`)、多选题选项 (`options`，仅适用于 `mcq` 类型)、附加提示 (`question_prompt`)、正确答案 (`answer`)、问题所使用图像的相对路径列表 (`image_paths`)，以及可选的辅助输入框信息 (`supporting.draw_boxes`)。数据集中的问题基于从同一静态场景的不同位置和角度捕获的多个视图图像，旨在评估模型的空间推理能力。该数据集主要用于视觉问答（VQA）任务的评估，特别侧重于多视图场景下的空间关系理解、物体检测和计数。
VIEW2SPACE v1 is a multi-view visual-language evaluation dataset for spatial reasoning. The dataset is associated with the ECCV 2026 paper titled VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations. The publicly released version includes three subsets: `count`, `detect`, and `mcq`. The data is organized in JSONL format, with each line in files such as `overall.jsonl` representing a question. Each data record contains the following fields: question ID (`q_idx`), question type (`q_type`), question text (`question`), multiple-choice options (`options`, applicable only to `mcq` type), additional prompt (`question_prompt`), correct answer (`answer`), list of relative paths to images used for the question (`image_paths`), and optional supporting input box information (`supporting.draw_boxes`). The questions in the dataset are based on multiple view images captured from different positions and angles of the same static scene, aiming to evaluate models spatial reasoning capabilities. The dataset is primarily used for evaluating visual question answering (VQA) tasks, with a particular focus on spatial relationship understanding, object detection, and counting in multi-view scenarios.
VIEW2SPACE v1 数据集概述
VIEW2SPACE v1 是一个面向空间推理的多视角视觉-语言评估数据集。该数据集包含了三个子集：count、detect 和 mcq。
基本信息
- 名称： VIEW2SPACE v1
- 任务类别： 视觉问答 (Visual Question Answering)
- 标签： 多视角、视觉-语言、空间推理、计数、检测、多选题
- 关联论文： VIEW2SPACE: Studying Multi-View Visual Reasoning from Sparse Observations（ECCV 2026）
- 论文地址： https://arxiv.org/abs/2603.16506
- 项目页面： https://pokerme7777.github.io/VIEW2SPACE/
- 许可证： Creative Commons Attribution 4.0 International License (CC BY 4.0)
相关发布
- 训练集： https://huggingface.co/datasets/Pokerme/view2space-train
- 4B 模型检查点： https://huggingface.co/Pokerme/view2space_4b
- 集合页面： https://huggingface.co/collections/Pokerme/view2space
子集与数据文件
数据集包含三个子集，每个子集对应一个测试分片 (test) 的 JSONL 文件：
| 配置名称 | 分片 | 文件路径 |
|---|---|---|
count |
test | count/overall.jsonl |
detect |
test | detect/overall.jsonl |
mcq |
test | mcq/overall.jsonl |
目录结构
view2space-v1-release/ README.md count/ overall.jsonl detect/ overall.jsonl mcq/ overall.jsonl images/ img_000001.png img_000002.png ...
数据格式
每条 JSONL 记录包含以下字段：
q_idx：问题 ID，例如mcq_000001。q_type：问题类别，值为mcq、detect或count。question：问题文本。options：多选题的选项；非多选题为空。question_prompt：附加提示文本。answer：真实答案。image_paths：问题所使用图片的相对路径列表（相对于数据集根目录，例如images/img_000123.png）。supporting.draw_boxes：可选的、与一张或多张图片关联的输入框。
示例记录
json { "q_idx": "mcq_000001", "q_type": "mcq", "question": "All views are captured from the same static scene arrangement, but from different positions and angles. In view 2, consider the nearest man with an orange hat, crouching. Where is the white temple relative to this man with an orange hat, crouching?", "options": { "A": "left", "B": "right", "C": "front", "D": "back" }, "question_prompt": "", "answer": "C", "image_paths": [ "images/img_000813.png", "images/img_000806.png", "images/img_000815.png" ], "supporting": { "draw_boxes": null } }
使用示例
python import json from pathlib import Path
dataset_root = Path("/path/to/view2space-v1-release") jsonl_path = dataset_root / "mcq" / "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"]] print(first["q_idx"]) print(first["q_type"]) print(image_files)
引用信息
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} }
