In the task of tooth segmentation, 3D intraoral scanning (IOS) technology has become an important data source due to its high resolution and convenience. However, the complex boundary morphology between teeth and gingiva makes it difficult for traditional segmentation methods to extract the tooth regions from oral data. This paper proposes a 3D tooth segmentation network based on geometric boundary enhancement and curvature guidance for the full set of teeth segmentation in IOS data. The main challenge of this task is not the subtle differences between teeth, but the accurate segmentation of the boundary between teeth and gingiva. To address this, we design a geometric boundary enhancement branch to improve the network’s sensitivity to the segmentation of tooth-gingiva boundary. To further enhance the network’s ability to segment the boundary, we introduce a curvature-guided loss that encourages the model to focus on the segmentation of the boundary, thereby reducing mis-segmentation of this region. Experimental results show that the proposed method achieves a segmentation accuracy of 96.63% on a test set containing 92 self-collected IOS data, outperforming the original Point Transformer and other mainstream methods. Moreover, the proposed method performs better under challenging conditions such as tooth missing, tooth crowding, and varying tooth counts. Ablation studies validate the effectiveness of the geometric boundary enhancement branch and the curvature-guided loss design.