A Spectral Keypoint Distillation Network for Point Cloud Registration

Deep learning-based point cloud registration often suffers from high computational costs, limiting deployment in resource-constrained applications. To address this, we propose Spectral Keypoint Distillation Registration (SKDR), a novel framework that creates a lightweight and efficient model via knowledge distillation. SKDR leverages spectral keypoints—stable geometric anchors derived from graph Laplacian analysis—to guide the transfer of structural insights from a high-performance teacher to a student network. By focusing on these critical points, the method ensures efficient geometric knowledge transfer. Experimental results demonstrate that SKDR achieves accuracy comparable to state-of-the-art heavy models while offering significantly faster inference speeds and outperforming other lightweight baselines in robustness.