omni-primitive-transforms
收藏资源简介:
Omni-Primitive-Transforms是一个用于生成受控3D物体原始变换序列的数据集。它基于高质量扫描网格(源自OmniObject3D数据集),并使用Blender进行渲染,旨在支持对单一或组合原始变换(如旋转、平移、缩放)下的3D物体序列进行可控生成。该数据集最初为世界模型和潜在动作模型的研究而创建。它包含5,911个日常物体实例,涵盖216个不同类别,每个实例提供72个渲染视图,对应物体绕垂直轴完整旋转360度(以5度为增量),所有图像分辨率为128×128像素。每个视图都配有对应的物体分割掩码(以NumPy数组格式存储),使得数据集能直接支持真实的3D旋转序列,并可通过2D掩码合成额外变换(如缩放和平移),从而支持可控变换序列生成。数据以分层目录结构组织:按物体类别和实例ID分文件夹存放,每个实例文件夹包含一个`Scan`子目录,存储从`000.png`到`071.png`的72张PNG格式渲染图像和对应的`NNN_mask.npy`分割掩码文件,以及一个`metadata.json`文件用于存储元数据。该数据集适用于需要可控3D物体变换序列的计算机视觉任务,特别是与世界模型、潜在动作建模、物体中心表示和原始变换学习相关的研究。渲染代码已开源,确保可复现性,使用需遵守OmniObject3D的许可条款。
Omni-Primitive-Transforms is a dataset for generating controlled 3D object primitive transformation sequences. It is built on high-quality scanned meshes sourced from the OmniObject3D dataset and rendered using Blender, with the aim of supporting controllable generation of 3D object sequences under single or combined primitive transformations such as rotation, translation, and scaling. This dataset was initially developed for research on world models and latent action models. It contains 5,911 everyday object instances spanning 216 distinct categories. Each instance provides 72 rendered views, corresponding to a full 360-degree rotation of the object around the vertical axis in 5-degree increments, with all images having a resolution of 128×128 pixels. Each view is paired with a corresponding object segmentation mask stored in NumPy array format, enabling the dataset to directly support realistic 3D rotation sequences. Additional transformations such as scaling and translation can also be synthesized using the 2D masks, thereby supporting controllable transformation sequence generation. The dataset is organized in a hierarchical directory structure: folders are categorized by object category and instance ID. Each instance folder contains a `Scan` subdirectory, which stores 72 PNG-format rendered images from `000.png` to `071.png`, the corresponding `NNN_mask.npy` segmentation mask files, as well as a `metadata.json` file for storing metadata. This dataset is applicable to computer vision tasks requiring controlled 3D object transformation sequences, particularly research related to world models, latent action modeling, object-centric representations, and primitive transformation learning. The rendering code has been open-sourced to ensure reproducibility, and its usage must comply with the license terms of OmniObject3D.




