OCMG-Net: Neural Oriented Normal Refinement for Unstructured Point Clouds

TL;DR

This work tackles oriented normal estimation on unstructured point clouds by introducing OCMG-Net, a soft refinement framework that starts from PCA+MST initialization and iteratively refines normals using a dual-parallel architecture for unoriented normals and orientation signs. A central contribution is the Chamfer Normal Distance (CND), defined as $\text{CND}(\mathcal{P}, \tilde{\mathcal{P}}) = \sqrt{\frac{1}{N} \sum_{i=1}^N \mathrm{arccos}^2 \langle \boldsymbol{n}_{\tilde{\boldsymbol{p}}_i}, \hat{\boldsymbol{n}}_{\boldsymbol{p}_i} \rangle}$, which aligns training with the geometry of the underlying surface under noise. The network employs Multi-scale Local Feature Aggregation and Hierarchical Geometric Information Fusion to capture complex geometry while mitigating scale ambiguity, and augments the orientation sign refinement with feature-space data augmentation and colorfully designed losses (including L1/L2/L3/L4/L5 components) that leverage CND. Extensive experiments across synthetic and real-world indoor/outdoor datasets show that OCMG-Net achieves state-of-the-art performance for both unoriented and oriented normals, improves downstream tasks like surface reconstruction and denoising, and generalizes well to unseen data. The method balances accuracy and efficiency, providing a practical, robust tool for point-cloud geometry processing with publicly available code.

Abstract

We present a robust refinement method for estimating oriented normals from unstructured point clouds. In contrast to previous approaches that either suffer from high computational complexity or fail to achieve desirable accuracy, our novel framework incorporates sign orientation and data augmentation in the feature space to refine the initial oriented normals, striking a balance between efficiency and accuracy. To address the issue of noise-caused direction inconsistency existing in previous approaches, we introduce a new metric called the Chamfer Normal Distance, which faithfully minimizes the estimation error by correcting the annotated normal with the closest point found on the potentially clean point cloud. This metric not only tackles the challenge but also aids in network training and significantly enhances network robustness against noise. Moreover, we propose an innovative dual-parallel architecture that integrates Multi-scale Local Feature Aggregation and Hierarchical Geometric Information Fusion, which enables the network to capture intricate geometric details more effectively and notably reduces ambiguity in scale selection. Extensive experiments demonstrate the superiority and versatility of our method in both unoriented and oriented normal estimation tasks across synthetic and real-world datasets among indoor and outdoor scenarios. The code is available at https://github.com/YingruiWoo/OCMG-Net.git.

Figures (16)

• Figure 1: The 2D schematic representation of the four principal oriented normal estimation manners.
• Figure 2: (a) A comparison between the annotated and the proposed CND-modified ground-truth (GT) normals, where the latter is more consistent with the underlying surface geometries. (b) Our method outperforms competitors with higher robustness to noise and intricate shape details, as highlighted by the heat map.
• Figure 3: (a) The annotated ground-truth normal $\boldsymbol{n}_{\boldsymbol{p}_i}$ of the noisy point $\boldsymbol{p}_i$ determined before noisy disturbance indeed is inconsistent with the input patch. (b) The direction of the normal $\boldsymbol{n}_{\tilde{\boldsymbol{p}}_i}$ of the nearest clean point $\tilde{\boldsymbol{p}}_i$ is more consistent with the input patch, which is taken as the CND-modified GT normal. (c) The predicted offset $\hat{\boldsymbol{d}}_{\boldsymbol{p}_i}$ cannot drag $\boldsymbol{p}_i$ to the noise-free underlying surface. (d) This inconsistency also degrades the surface reconstruction performance.
• Figure 4: The schematic pipeline of the proposed OCMG-Net for orientated normal estimation. The initialized oriented normals acquired by PCA and MST are refined in the Refinement Module. This refinement process leverages features extracted from the dual-parallel Multi-scale Local Feature Extraction and Hierarchical Geometric Information Fusion approaches. Through this process, our method consistently delivers high-quality oriented normals, thereby ensuring accurate surface reconstruction.
• Figure 5: The architecture of the proposed Multi-scale Local Feature Aggregation module.