Accurately delineating individual teeth in 3-dimensional tooth point clouds is an important orthodontic application. Learning-based segmentation methods rely on labeled datasets, which are typically limited in scale due to the labor-intensive process of annotating each tooth. In this article, we propose a self-supervised pretraining framework, named Geo-Net, to boost segmentation performance by leveraging large-scale unlabeled data. The framework is based on the scalable masked autoencoders, and 2 geometry-guided designs, curvature-aware patching algorithm (CPA) and scale-aware reconstruction (SCR), are proposed to enhance the masked pretraining for tooth point cloud segmentation. In particular, CPA is designed to assemble informative patches as the reconstruction unit, guided by the estimated pointwise curvatures. Aimed at equipping the pretrained encoder with scale-aware modeling capacity, we also propose SCR to perform multiple reconstructions across shallow and deep layers. In vitro experiments reveal that after pretraining with large-scale unlabeled data, the proposed Geo-Net can outperform the supervised counterparts in mean Intersection of Union (mIoU) with the same amount of annotated labeled data. The code and data are available at https://github.com/yifliu3/Geo-Net .

中文翻译：

---

Geo-Net：用于齿点云分割的几何引导预训练


在 3 维牙齿点云中准确描绘单颗牙齿是一项重要的正畸应用。基于学习的分割方法依赖于标记的数据集，由于注释每颗牙齿的过程需要劳动密集型，因此数据集的规模通常受到限制。在本文中，我们提出了一个名为 Geo-Net 的自我监督预训练框架，通过利用大规模未标记数据来提高分割性能。该框架基于可扩展的掩码自动编码器，并提出了 2 种几何导向设计，即曲率感知修补算法 （CPA） 和尺度感知重建 （SCR），以增强齿点云分割的掩码预训练。特别是，CPA 旨在组装信息斑块作为重建单元，以估计的逐点曲率为指导。为了使预训练的编码器具有尺度感知建模能力，我们还提出了 SCR 在浅层和深层之间执行多次重建。体外实验表明，在用大规模未标记数据进行预训练后，在相同数量的注释标记数据下，所提出的 Geo-Net 在平均联合交集 （mIoU） 方面优于监督的对应物。代码和数据可在 https://github.com/yifliu3/Geo-Net 上获得。