ffp
收藏资源简介:
FoundationPose Paired Synthetic Renders数据集是一个用于3D视觉任务的合成渲染数据集,包含来自GSO(Google Scanned Objects)和Objaverse的配对视图。数据集分为四个配置:gso-scenes、gso-assets、objaverse-scenes和objaverse-assets,其中场景配置包含训练和验证分割,资产配置仅包含训练分割。每个场景行提供两个同步视图,并标注了每个对象的姿态、掩码ID、边界框、可见性和遮挡信息。数据以多种格式存储:深度数据为float32 NPY字节,RGB图像和uint32实例掩码为PNG字节。矩阵使用行主序扁平化数组表示,并采用规范的列向量名称(如camera_from_world和world_from_object),同时保留了原始源变换。资产表仅包含稳定的源标识符,渲染档案引用外部USD几何体,不包含可重用的网格资产。该数据集适用于对象姿态估计、场景理解和计算机视觉研究,支持基于合成渲染数据的模型训练和评估。
The FoundationPose Paired Synthetic Renders dataset is a synthetic rendering dataset for 3D vision tasks, containing paired views sourced from GSO (Google Scanned Objects) and Objaverse. The dataset is divided into four configurations: gso-scenes, gso-assets, objaverse-scenes, and objaverse-assets. Among these, the scene configurations include training and validation splits, while the asset configurations only contain training splits. Each scene entry provides two synchronized views, with annotations including pose, mask ID, bounding box, visibility, and occlusion information for each object. Data is stored in multiple formats: depth data is stored as float32 NPY binary files, while RGB images and uint32 instance masks are stored as PNG-encoded binary data. Matrices are represented as row-major flattened arrays, following standard column-vector naming conventions (e.g., camera_from_world and world_from_object), while retaining the original source transformations. The asset tables only contain stable source identifiers, and the render archives reference external USD geometry without including reusable mesh assets. This dataset is applicable to object pose estimation, scene understanding, and computer vision research, supporting model training and evaluation based on synthetic rendering data.




