VP inconsistencies dataset
收藏资源简介:
该数据集由东京大学研究团队创建,专注于AI生成图像中的消失点(VP)不一致性问题,旨在为几何校正提供量化评估基准。数据集包含人工标注的消失点位置和建筑轮廓,数据来源为Stable Diffusion等主流生成模型产生的图像,并通过专业标注构建几何一致性真值。研究团队采用交互式标注流程,确保数据精准反映建筑场景中平行线组的空间几何关系。该数据集主要应用于计算机视觉领域,解决生成图像的结构真实性缺陷,并为图像到3D重建等需几何精确度的任务提供支持。
This dataset was developed by a research team at The University of Tokyo, focusing on the vanishing point (VP) inconsistency issue in AI-generated images, and aims to provide a quantitative evaluation benchmark for geometric correction. It contains manually annotated vanishing point positions and building outlines. The source images are generated by mainstream generative models including Stable Diffusion, and professional annotation is employed to establish the ground truth of geometric consistency. The research team utilized an interactive annotation workflow to ensure that the data accurately reflects the spatial geometric relationships of parallel line groups in architectural scenes. This dataset is primarily applied in the field of computer vision, addressing structural authenticity defects in generated images and supporting tasks requiring high geometric precision such as image-to-3D reconstruction.
ControlVP 数据集概述
数据集基本信息
- 数据集名称: HoliCity
- 关联论文: ControlVP: Interactive Geometric Refinement of AI-Generated Images with Consistent Vanishing Points
- 论文作者: Ryota Okumura, Kaede Shiohara, Toshihiko Yamasaki
- 发表会议: WACV 2026
- 论文链接: https://arxiv.org/abs/2512.07504
数据集用途
该数据集用于训练 ControlVP 模型,这是一个用户引导的框架,旨在利用建筑轮廓作为条件,校正 AI 生成图像中的消失点不一致问题。
数据集内容与结构
数据集包含以下文件,需从 HoliCity 主页 (https://holicity.io/) 下载:
- 图像文件: https://huggingface.co/yichaozhou/holicity/tree/main/perspective/image-v1
- 法线图: https://huggingface.co/yichaozhou/holicity/tree/main/perspective/normal-v1
- 平面图: https://huggingface.co/yichaozhou/holicity/blob/main/perspective/plane-v1.tar
- 消失点数据: https://huggingface.co/yichaozhou/holicity/blob/main/perspective/vpts-v1.tar
下载后,需按以下目录结构组织数据集:
<dataset_root>/ ├── images/ ├── normal_map/ ├── planes/ └── vanishing_points/
数据集预处理
需运行脚本 src/script/create_training_dataset.sh 来创建训练数据集。此过程将生成包含边缘信息的输出目录。
相关资源
- 官方代码仓库: https://github.com/RyotaOkumura/ControlVP
- 交互式 GUI 工具: https://github.com/RyotaOkumura/GUI_for_ControlVP (用于创建自定义控制指令)
- 预训练模型: 代码仓库提供,可用于快速推理。

- 1ControlVP: Interactive Geometric Refinement of AI-Generated Images with Consistent Vanishing Points东京大学 · 2025年



