Accurate registration of intraoral scans (IOS) and cone-beam computed tomography (CBCT) is a critical prerequisite for precise diagnosis and treatment planning in dentistry. However, large modality discrepancies and dense point clouds make this task challenging in practice. In this work, we propose a learning-based framework for CBCT–IOS registration, developed in the context of the MICCAI STSR Task 2 2025 Challenge. Our method leverages dual PointNet++ encoders to extract modality-specific features, followed by a differentiable SVD head that execute rigid-body constraints in the predicted transformation. To enhance robustness, we design geometric data augmentation strategies, while point cloud sampling and simplification are employed to accelerate inference. Ablation studies demonstrate that augmentation substantially reduces registration errors, while relaxing CBCT filtering thresholds further improves alignment by preserving richer anatomical cues. Overall, our approach achieves competitive performance, ranking second on the validation leaderboard, and provides a practical balance between accuracy and efficiency.