Symmetria
收藏资源简介:
Symmetria是一个基于公式的数据集,可以生成任意规模的点云数据。它通过构建确保精确的地面真值可用,促进数据高效实验,实现多样化几何设置下的广泛泛化,并易于扩展到新任务和模式。使用对称性的概念,我们创建具有已知结构和高度可变性的形状,使神经网络能够有效地学习点云特征。我们的结果表明,这个数据集对于点云自监督预训练非常有效,产生了在下游任务(如分类和分割)中表现强劲的模型,这些模型也显示出良好的少样本学习能力。此外,我们的数据集可以支持对现实世界对象进行微调的模型分类,突出了我们的方法在实际应用中的实用性和应用价值。我们还引入了对称性检测的挑战性任务,并为基线比较提供了一个基准。我们方法的一个显著优势是数据集的公共可用性、伴随代码,以及能够生成非常大的集合,这促进了点云领域进一步的研究和创新。
Symmetria is a formula-based dataset that can generate point cloud data of arbitrary scale. It ensures precise ground truth availability through its construction, facilitating data-efficient experiments, enabling broad generalization across diverse geometric configurations, and allowing straightforward extension to new tasks and modalities. Drawing on the concept of symmetry, we create shapes with well-defined structures and high variability, enabling neural networks to effectively learn point cloud features. Our results show that this dataset is highly effective for self-supervised pre-training of point clouds, yielding models that achieve strong performance on downstream tasks such as classification and segmentation, and which also exhibit excellent few-shot learning capabilities. Furthermore, our dataset supports classification tasks for models fine-tuned on real-world objects, highlighting the practical utility and application value of our approach in real-world scenarios. We also introduce a challenging symmetry detection task and provide a benchmark for baseline comparisons. One notable advantage of our method is the public availability of the dataset, the accompanying code, and the ability to generate extremely large-scale collections, which advances further research and innovation in the point cloud domain.




