uva-cv-lab/Point-MAE-Zero

Procedural 3D Synthetic Shapes数据集包含152,508个通过程序合成的3D形状，旨在帮助人们更好地复现《Learning 3D Representations from Procedural 3D Programs》论文中的结果。这些形状通过结合基本形状（如立方体、球体和圆柱体）并应用各种变换和增强来创建，以增强几何多样性。每个3D形状由均匀采样的表面点表示，适用于预训练模型，用于任务如掩码点云补全、形状分类等。数据集格式为点云学习准备，存储为.npy文件，包含152,508个子目录，每个目录中提供object.npy和object_aug.npy文件。

The Procedural 3D Synthetic Shapes dataset contains 152,508 procedurally synthesized 3D shapes to help people better reproduce results for the paper Learning 3D Representations from Procedural 3D Programs. The shapes are created by combining primitive shapes (e.g., cubes, spheres, and cylinders) and applying various transformations and augmentations to enhance geometric diversity. Each 3D shape is represented by uniformly sampled surface points, making it suitable for pretraining models for tasks such as masked point cloud completion, shape classification, and more. The dataset is provided in a format ready for point cloud-based learning, stored as .npy files, with 152,508 sub-directories, each containing object.npy and object_aug.npy files.