Accurate segmentation of teeth in 3D intraoral scanning data is crucial for diagnosis and treatment planning. To address the pain points of existing algorithms, such as low accuracy in classification and segmentation as well as high memory consumption when dealing with unclear tooth boundaries, crowding, misalignment, and missing teeth, this paper proposes a two-stage tooth classification and segmentation model that uses an attention mechanism to enhance feature extraction. In the first stage, 3D point clouds are first projected into 2D images. Aiming at the problem of low tooth classification accuracy in existing methods, on the basis of the YOLOv5x algorithm, a tooth Feature Enhancement Learning Module (FELM) constructed by Transformer is added to strengthen the learning of correlations between data points in each tooth region, thereby improving the classification accuracy of detection. This results in the detection model YOLOv5x-FELM. In the second stage, after detecting the 2D tooth regions, it is necessary to segment the corresponding single-tooth point cloud data on the cropped 3D intraoral scanning model. To solve the problem that EdgeConv can only learn local features, a Global Channel Relation Feature (GCRF) module based on the attention mechanism is added to enhance the fusion and learning of global features, resulting in EdgeConv-GCRF. Compared with existing two-stage methods, the proposed method in this paper improves the accuracy of classification and segmentation. Experiments on the Teeth3DS dataset show that, compared with existing methods, this method achieves the highest mean accuracy (mAcc), mean Intersection over Union (mIOU), and mean Dice score (mDice): 0.8608, 0.8249, and 0.8666, respectively. The detection accuracy is significantly improved, and it can efficiently complete the segmentation task, which has great economic and social value.