hpotechius/IntrinsicDecompositionDataset

--- license: cc-by-nc-4.0 task_categories: - image-to-image tags: - color-transfer - indoor-scene --- # Intrinsic Decomposition Dataset The Interior Decomposition Dataset (IDD) is a dataset designed to analyze how intrinsic scene properties influence the perceived quality of color transfer methods. It consists of photorealistic, synthetically rendered indoor scenes with systematically controlled variations in color, illumination, geometry, object arrangement, and viewpoint. In addition to the rendered images, the dataset provides detailed ground-truth information such as semantic segmentation, depth, and physically-based rendering passes. ## Dataset Viewer [](https://potechius.com/IDD-Viewer/) ## Details - The dataset consists of 4 photorealistic indoor scenes with systematically controlled variations. - Each scene is provided in 4 different color configurations, 8 distinct illumination settings, 4 different object arrangements, and 6 camera viewpoints. - In total, the dataset contains 3,072 unique scene instances. - All images are stored at a resolution of 1024 × 1024 pixels in 32-bit floating-point EXR format using a linear RGB color space. - Each scene instance includes a semantic map with 26 channels representing object classes using soft assignments. - Depth maps are provided with values ranging from 0 to 15 meters, normalized to the interval [0, 1]. - Normal maps are available in both world-space and view-space representations. - Pixel size maps encode the estimated real-world surface area covered by each pixel based on depth and viewing angle. - A beauty pass decomposition is included, consisting of 14 rendering channels such as diffuse and reflection components. - The rendered images are released under a Creative Commons CC BY-NC 4.0 license, while the original 3D models are not included. <p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/6331b00ab7b8e9d6e89a5dc0/c97BaDEU9T5GjiCg34y2D.webp" width="512"> <em>Figure 1: Scene variations</em> </p> <p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/6331b00ab7b8e9d6e89a5dc0/K7xfumn59VtpBzaobiTO3.webp" width="512"> <em>Figure 2: Extract from the IDD: base scene, with variations in color statistics (top left to right), illumination, object arrangement, and view (bottom left to right).</em> </p> <p align="center"> <img src="https://cdn-uploads.huggingface.co/production/uploads/6331b00ab7b8e9d6e89a5dc0/Ux7UyRVTY3MMBWJBIt8Z_.webp" width="512"> <em>Figure 3: Illustration of scene cues for color transfer: semantic map, size map, diffuse pass (left to right).</em> </p> ## Citation If you utilize this dataset in your research, kindly provide a citation: ``` @inproceedings{potechius2026, author={Potechius, H. and Sikora, T. and Knorr, S.}, booktitle={IEEE International Conference on Image Processing (ICIP)}, title={{The Impact of Intrinsic Scene Cues on Perceived Color Transfer Quality}}, year={2026}, location = {Tampere, Finland} } ```The Interior Decomposition Dataset (IDD) is a dataset designed to analyze how intrinsic scene properties influence the perceived quality of color transfer methods. It consists of photorealistic, synthetically rendered indoor scenes with systematically controlled variations in color, illumination, geometry, object arrangement, and viewpoint. In addition to the rendered images, the dataset provides detailed ground-truth information such as semantic segmentation, depth, and physically-based rendering passes. Specifically, the dataset consists of 4 photorealistic indoor scenes with systematically controlled variations. Each scene is provided in 4 different color configurations, 8 distinct illumination settings, 4 different object arrangements, and 6 camera viewpoints, totaling 3,072 unique scene instances. All images are stored at a resolution of 1024 × 1024 pixels in 32-bit floating-point EXR format using a linear RGB color space. Each scene instance includes a semantic map with 26 channels representing object classes using soft assignments, depth maps with values ranging from 0 to 15 meters normalized to the interval [0, 1], normal maps in both world-space and view-space representations, pixel size maps encoding the estimated real-world surface area covered by each pixel based on depth and viewing angle, and a beauty pass decomposition consisting of 14 rendering channels such as diffuse and reflection components. The rendered images are released under a Creative Commons CC BY-NC 4.0 license, while the original 3D models are not included.

搜集汇总

数据集介绍

构建方式

IntrinsicDecompositionDataset旨在探究内在场景属性对色彩迁移方法感知质量的影响，其构建基于四张逼真的室内场景合成渲染图。通过系统化控制色彩配置、照明条件、物体排列与相机视角等变量，每张场景衍生出4种色彩方案、8种光照、4种布局及6个视点，最终生成3072个独特场景实例。所有图像以32位浮点EXR格式存储于线性RGB色彩空间，分辨率统一为1024×1024像素，并附带丰富的真实标注信息，包括语义分割图、深度图、法线图、像素尺寸图以及由14个渲染通道构成的美学通道分解。

特点

该数据集的核心特点在于其多维度的系统变异性与高精度标注的有机结合。场景变量覆盖色彩、光照、几何、物体排列与视点五大维度，为分析内在场景线索对色彩迁移的独立或交互影响提供了可控实验环境。除常规图像外，数据集还提供了26通道语义图（软分配）、归一化至[0,1]的深度图、世界空间与视图空间法线图、基于深度与视角估算的像素尺寸图，以及包含漫反射与反射成分的14通道渲染分解，为计算机视觉中的场景理解与图像处理研究提供了丰富的基础数据资源。

使用方法

用户可通过HuggingFace平台直接访问数据集，利用提供的IDD Viewer工具浏览场景变体及对应标注。数据以EXR格式存储，支持主流图像处理库（如OpenEXR、PyTorch或TensorFlow）进行加载与分析。研究实践中，可选取特定变量组合的子集用于色彩迁移算法评估，或利用语义、深度、法线等多模态标注训练场景理解模型。需注意数据集遵循CC BY-NC 4.0许可证，原始3D模型未包含在内，使用时须正确引用原始文献。

背景与挑战

背景概述

IntrinsicDecompositionDataset（亦称Interior Decomposition Dataset，IDD）是由Potechius、Sikora和Knorr等人于2026年在IEEE International Conference on Image Processing（ICIP）上提出的一个合成室内场景数据集。该数据集旨在系统性地探究内在场景属性（如颜色、光照、几何结构、物体布局与视角）对色彩迁移方法感知质量的影响。通过提供4个逼真室内场景在4种颜色配置、8种光照条件、4种物体排列及6种视角下的3072个独有实例，IDD填补了色彩迁移领域缺乏受控对照数据的空白。每张1024×1024像素的32位浮点EXR图像均包含语义分割、深度图、法线图、像素尺寸图及14通道渲染分解等详细真值信息，为理解场景固有特性与图像感知质量之间的关联提供了宝贵资源。该数据集现以CC BY-NC 4.0许可发布，在计算机视觉与图形学领域具有重要参考价值，尤其推动了色彩迁移、内在图像分解及场景理解等方向的研究进展。

当前挑战

该数据集所解决的领域问题主要聚焦于色彩迁移方法对场景内在属性的依赖机制，传统色彩迁移研究多局限于全局颜色统计或简单对应关系，而IDD揭示了光照、几何、材质等复杂因素对迁移效果的影响，为开发鲁棒性更强的算法提供了高通真度的基准。在构建过程中，挑战在于合成场景的逼真度与变量控制之间的平衡：需确保渲染图像具备照片级真实感的同时，独立调控颜色、光照、物体布局等维度；此外，为每张图像生成完整的语义标注、深度及多通道渲染分解（如漫反射与反射分量）涉及大量计算资源与精确的物理模拟；同时，32位浮点EXR格式的存储与标准化处理也对数据管线提出了严格的技术要求。数据集以CC BY-NC 4.0许可发布，但原始3D模型未包含在内，可能限制了场景扩展与二次开发的灵活性。

常用场景

经典使用场景

在计算机视觉与图形学领域，IntrinsicDecompositionDataset被广泛用于探究色彩迁移技术的内在机理。研究人员可借助该数据集系统性地操控室内场景的颜色、光照、几何结构、物体布局及视角等多元属性，从而量化分析这些固有场景线索如何影响色彩迁移结果的主观感知质量。通过对比不同场景配置下的色彩迁移效果，该数据集为理解场景内在因素与视觉质量之间的复杂关联提供了扎实的基准。

实际应用

在实际应用中，该数据集为图像编辑软件、影视后期制作、增强现实等视觉内容创作工具的质量评估与算法优化提供了重要支撑。开发人员可利用其多样化的场景实例测试和调校色彩迁移算法的鲁棒性与适应性，确保在复杂真实光照和格局下仍能保持视觉和谐。此外，数据集中的语义分割与深度信息还可服务于智能图像理解系统，提升自动色彩校正与风格迁移功能的精准度。

衍生相关工作

围绕该数据集衍生出一系列经典工作，包括构建基于场景线索的感知质量预测模型、开发面向精细色彩传递的新型深度学习框架，以及探索多模态真值（如反射分量、法线图）在图像分解任务中的协同作用。相关研究不仅在ICIP等顶级会议上发表，还催生了诸如色彩迁移主观评价基准、光照鲁棒性评估协议等工具，进一步拓宽了该数据集在视觉质量工程领域的学术影响力。