ADSKAILab/Zero-To-CAD-100k
收藏资源简介:
这是Zero-to-CAD的精选100K子集,旨在为计算资源有限的研究者提供一个易于访问的入口。样本从完整的1M数据集中选出,以确保最大的几何多样性。通过视觉嵌入(使用DINOv3特征)、聚类(K-means)和选择(每个簇中选择最接近中心的样本)的过程,确保子集涵盖了1M数据集中所有零件类型、操作和几何复杂性的完整分布。数据集包含81,015个训练样本、9,734个验证样本和9,767个测试样本,每个样本包含唯一标识符、可执行的CadQuery Python源代码、B-Rep面数、CAD操作列表、渲染视图、STL和STEP文件等字段。
This is a curated 100K subset of Zero-to-CAD, designed to provide an easily accessible entry point for researchers with limited computational resources. Samples are selected from the full 1M dataset to ensure maximum geometric diversity. Through a pipeline of visual embedding (using DINOv3 features), K-means clustering, and sample selection (selecting the sample closest to the cluster centroid in each cluster), the subset is guaranteed to cover the full distribution of all part types, operations, and geometric complexities present in the 1M dataset. The dataset contains 81,015 training samples, 9,734 validation samples, and 9,767 test samples. Each sample includes fields such as unique identifier, executable CadQuery Python source code, number of B-Rep faces, list of CAD operations, rendered views, STL files, and STEP files.




