NCVP-Reg: Point Cloud Registration via Non-local Feature Fusion and Uncertainty Calibrated Virtual Points

Facial point cloud registration is essential for minimally invasive head and neck surgery, enabling accurate alignment between virtual anatomical models and patient facial tissues. However, surgical scenarios often require registration between partial and complete point clouds, where differences in scale and structure make it difficult to achieve both robustness and efficiency. To address this challenge, we propose NCVP-Reg, a point cloud registration framework based on Non-local Feature Fusion and Uncertainty Calibrated Virtual Points. A Non-local Branch Fusion module integrates global contextual information into local feature encoding, improving robustness to scale variations. Meanwhile, Uncertainty Calibrated Virtual Points (UCVP) generate adaptive virtual correspondences to mitigate structural discrepancies and enhance part-to-whole registration. Experiments demonstrate the effectiveness of the proposed method. Experiments demonstrate the effectiveness of the proposed method. On the BIT Face3D dataset with simulated facial occlusions, NCVP-Reg achieves a mean pairwise distance error of 2.32 mm and a mean recall ratio of 99.63% at 71.06 FPS. On the BARIM dataset with diverse head poses, it maintains a mean 96.65% recall ratio. Under the unseen-category setting on ModelNet40, the method achieves a relative rotation error of 1.62◦, demonstrating strong generalization.