Visual Causal Graph dataset (VCG-32K)
收藏资源简介:
VCG-32K是由上海人工智能实验室等机构联合创建的大规模视觉因果图数据集,旨在推动视觉因果发现研究。该数据集包含32,256张图像,涵盖299,262个实体、2,287个实体类别和185,321条因果关系,数据源自MS-COCO和Objects365两大经典视觉理解数据集。其构建过程采用两阶段标注流程,首先精修边界框以确保实体定位准确,随后基于直接接触、状态维持与反事实干预等因果原则标注实体间的因果机制类型。该数据集主要应用于训练和评估视觉语言模型进行因果推理,以解决机器人操作、自动驾驶等需要安全可靠决策的下游任务中,模型缺乏深层因果理解能力的核心挑战。
VCG-32K is a large-scale visual causal graph dataset jointly created by Shanghai AI Laboratory and other institutions, aiming to advance research in visual causal discovery. This dataset contains 32,256 images, covering 299,262 entities, 2,287 entity categories and 185,321 causal relationships, and its data is sourced from two classic visual understanding datasets: MS-COCO and Objects365. Its construction adopts a two-stage annotation pipeline: first, bounding boxes are refined to guarantee accurate entity localization, then the types of causal mechanisms between entities are annotated based on causal principles including direct contact, state persistence and counterfactual intervention. This dataset is primarily used for training and evaluating visual language models to conduct causal reasoning, so as to address the core challenge that models lack deep causal understanding capabilities in downstream tasks requiring safe and reliable decision-making such as robotic manipulation and autonomous driving.
CauSight 数据集概述
数据集基本信息
- 数据集名称:VCG-32K
- 数据集发布者:OpenCausaLab
- 数据集地址:https://huggingface.co/datasets/OpenCausaLab/VCG-32K
- 关联模型:CauSight
- 关联模型地址:https://huggingface.co/OpenCausaLab/CauSight
- 关联论文:https://arxiv.org/abs/2512.01827
- 论文标题:CauSight: Learning to Supersense for Visual Causal Discovery
数据集内容与用途
该数据集用于视觉因果发现任务,旨在支持模型学习“超感知”以进行视觉因果推理。
数据集获取与使用
下载方式
通过Hugging Face Hub下载,需使用huggingface_hub库。
bash
hf download OpenCausaLab/VCG-32K --repo-type dataset --local-dir ./VCG-32K
数据预处理
下载后包含压缩文件,需解压至指定目录。
- COCO图像数据:
./VCG-32K/COCO/images.tar.gz - 365图像数据:
./VCG-32K/365/images.tar.gz
关联资源
- 代码仓库:https://github.com/OpenCausaLab/CauSight
- 环境配置:需使用Python 3.10,依赖包见
requirements.txt。 - 评估流程:需启动模型服务器后运行推理脚本。
- 扩展功能:支持使用Tree-of-Causal-Thought方法生成自定义SFT数据。

- 1CauSight: Learning to Supersense for Visual Causal Discovery上海人工智能实验室、上海创新研究院、上海交通大学、北京大学、同济大学 · 2025年



