Abstract
Point cloud filtering is a fundamental problem in geometry modeling and processing. Despite significant advancement in recent years, existing methods still suffer from two issues: 1) they are either designed without preserving sharp features or are less robust in feature preservation; and 2) they usually have many parameters and require tedious parameter tuning. In this paper, we propose a novel deep learning approach that automatically and robustly filters point clouds by removing noise and preserving their sharp features. Our point-wise learning architecture consists of an encoder and a decoder. The encoder directly takes points (a point and its neighbors) as input and learns a latent representation vector which goes through the decoder to relate the ground-truth position with a displacement vector. The trained neural network can automatically generate a set of clean points from a noisy input. Extensive experiments show that our approach outperforms state-of-the-art deep learning techniques in terms of both visual quality and quantitative error metrics.
Introduction
3D scanning, whether from consumer-grade depth sensors or high-end LIDAR, inevitably produces noisy point clouds. To utilize this data for downstream tasks like 3D printing, autonomous driving, or digital archiving, the noise must be filtered out. The primary challenge lies in the "smoothing vs. sharpening" trade-off. Traditional algorithms, such as LOP or RIMLS, often require users to manually tune complex parameters; incorrect settings can either leave noise behind or "melt" sharp edges. Early deep learning attempts struggled with generalization or required manual labeling of sharp edges. Pointfilter addresses these issues by providing an automatic, robust method capable of keeping edges crisp without human intervention.
Methodology
Pointfilter treats denoising as a regression problem using a point-wise learning architecture that focuses on local geometry patches.
Encoder-Decoder Architecture
The network processes a noisy point and its local neighbors. The Encoder, based on the PointNet architecture, extracts a latent representation of the local geometry. The Decoder then regresses a displacement vector—a "push" that moves the noisy point back to its rightful place on the underlying clean surface.
PCA Alignment and Loss Functions
To ensure rotation invariance, the authors use Principal Component Analysis (PCA) to align each local patch to a canonical coordinate system. The core innovation is the bilateral projection loss ($L^b_{proj}$). Unlike standard $L_2$ loss, this incorporates normal similarity during training to ensure points are projected onto correct surface patches, preventing the blurring of sharp corners. Additionally, a repulsion term ($L_{rep}$) encourages a uniform distribution of points to prevent clumping.
Experimental Results
The model was trained on a curated dataset of 22 3D models (CAD and organic) augmented with synthetic Gaussian noise. Evaluation was performed on 15 unseen synthetic models and real-world scans from Kinect and LIDAR sensors.
Quantitative and Visual Performance
Pointfilter achieved the lowest errors across major metrics: Chamfer Distance (CD), Mean Square Error (MSE), and Point-to-Surface (P2F) distance. Notably, its CD was approximately 12% lower than the next best deep learning method. Qualitatively, Pointfilter excels at preserving 90-degree angles and fine ridges where other methods tend to create rounded "fillets" or leave outliers.
Efficiency and Practicality
While not the absolute fastest, Pointfilter is roughly 6x faster than competing methods like PCN and eliminates the "trial-and-error" time sink of manual tuning. A key practical advantage is that while ground-truth normals are required for training, they are not needed during inference, making it suitable for raw scan data where reliable normals are unavailable.
Conclusion
Pointfilter provides a robust, automatic solution for point cloud denoising with a focus on sharp feature preservation. While it faces challenges with extreme noise levels and large holes due to its local-patch focus, future work aims to incorporate global shape information to address these limitations.
Summary updated