Clean-Prior Distillation and Noise-Adaptive Structural Restoration for Point Cloud Denoising

Point cloud denoising must remove measurement noise without erasing fine geometric structure. Structural priors extracted from noisy features may mix these signals, while uniform feature fusion and coordinate updates may move points that are already reliable. We address the two problems separately. Hierarchical Geometry-Aware Clean-Prior Distillation (HGCPD) transfers clean structural knowledge from a gradient-frozen teacher, updated by an exponential moving average, to a masked noisy student. Its targets include clean-code assignments and multi-scale feature and geometry cues. Noise-Adaptive Structural Restoration (NASR) regulates decoder skip features and pointwise residuals through a Noise-Aware Structural Refinement Module (NSRM) and an Update-Gated Residual Head (UGRH). The teacher, masked branch, and clean codebook are used only during training, so inference retains the complete student restoration path. This combination addresses over- and under-denoising while preserving geometric detail. Across benchmarks, point densities, Gaussian noise levels, and unseen noise distributions, the model achieves lower mean denoising errors, with the clearest gains under dense sampling and strong corruption. Component studies indicate that clean-prior distillation and noise-adaptive restoration make complementary contributions. Project page: https://nasr-hgcpd.github.io/.