A Learning-Free Noise-Adaptive Framework for Feature-Preserving Point Cloud Denoising

Point cloud denoising is a fundamental preprocessing task in 3D vision and geometry processing, where the main challenge is to suppress corruption while preserving sharp features, thin structures, and local surface fidelity. Classical geometric filters are computationally efficient and interpretable, but they commonly rely on fixed local supports or globally selected parameters, which limits their effectiveness under spatially heterogeneous corruption. More adaptive non-local and learning-based approaches can improve robustness, yet they often introduce higher computational complexity, stronger modeling assumptions, or substantial training-data dependency. In this work, we propose Noise-Adaptive Bilateral Normal Projection (NABNP), a learning-free point cloud denoising framework that introduces explicit patch-wise adaptation to local corruption conditions. NABNP estimates a robust dimensionless local noise level from point-to-plane residuals and uses this quantity to adapt the angular bandwidth of bilateral normal refinement, the balance between weighted local averaging and normal projection, and the magnitude of the positional update. This design enables conservative smoothing in locally reliable neighborhoods while applying stronger geometric correction in more severely corrupted regions. We evaluate NABNP on standard benchmark models under 13 stress scenarios covering additive Gaussian noise, outlier contamination, sparsity, and rotation perturbation, resulting in 650 trials per method. The experimental results show that NABNP provides strong aggregate behavior among the evaluated learning-free baselines, particularly under low-to-medium corruption, sparsity, and rotation perturbations, while its advantage becomes less pronounced under the most severe noise and outlier settings. The method remains training-free and interpretable, with a moderate computational cost associated with repeated neighborhood analysis and covariance-based local updates.