Abstract
Establishing a consistent normal orientation for point clouds is a notoriously difficult problem in geometry processing, requiring attention to both local and global shape characteristics. The normal direction of a point is a function of the local surface neighborhood; yet, point clouds do not disclose the full underlying surface structure. Even assuming known geodesic proximity, calculating a consistent normal orientation requires the global context. In this work, we introduce a novel approach for establishing a globally consistent normal orientation for point clouds. Our solution separates the local and global components into two different sub-problems. In the local phase, we train a neural network to learn a coherent normal direction per patch (i.e., consistently oriented normals within a single patch). In the global phase, we propagate the orientation across all coherent patches using a dipole propagation. Our dipole propagation decides to orient each patch using the electric field defined by all previously orientated patches. This gives rise to a global propagation that is stable, as well as being robust to nearby surfaces, holes, sharp features and noise.
Introduction
In computer graphics and 3D vision, shapes are frequently captured as point clouds—unstructured sets of $(x, y, z)$ coordinates. Converting these points into a continuous surface requires accurate normal vectors at each point. Crucially, these normals must be oriented consistently, typically pointing "outward" from the object's interior. This task is challenging due to local ambiguities (where "up" and "down" look identical), the failure of greedy propagation methods in the presence of noise, and complex geometries like thin surfaces where algorithms may "short-circuit" across gaps.
We present a two-phase framework that decouples local consistency from global orientation. By leveraging a neural network for local patch coherency and a physics-inspired dipole propagation for global alignment, our method achieves state-of-the-art robustness and accuracy even in the presence of heavy noise and sharp features.
Methodology
Local Phase: Coherent Neural Patches
The point cloud is first partitioned into small, overlapping patches. We employ an ensemble of PointCNN architectures to ensure local coherency. The goal here is not to determine the absolute global orientation, but to ensure that all points within a single patch are oriented in the same direction (either all inward or all outward). The network also outputs a confidence score for each point, which is critical for the subsequent global phase.
Global Phase: Dipole Propagation
The core of our approach is the global propagation phase, which treats each oriented patch as an electric dipole.
Already-oriented patches generate a global electric field (a harmonic potential). To orient a new patch, we measure its interaction with this cumulative field. If the patch's current orientation conflicts with the field, it is flipped. Unlike traditional Minimum Spanning Tree (MST) methods that rely on local, greedy decisions, dipole propagation is informed by the entire set of previously oriented points. This makes the process highly stable and resistant to local noise or geometric complexities like sharp features.
Final Diffusion
Following the global propagation, a final diffusion step is performed. This step utilizes the global field to refine the orientation of individual points that may have been misclassified or missed during the initial neural patch phase.
Evaluation and Results
The model was trained on 1,410 patches derived from 60 watertight meshes (Thingi10K and non-rigid human scans), augmented with noise. Evaluation was conducted on diverse datasets, including large-scale scans from the Laric repository and point clouds generated by networks like EC-Net and PC-GAN.
Our method achieved an average of 96% correctly oriented normals, significantly outperforming the "No Net" baseline (74%) and traditional MST-based solvers like Hoppe’s or König’s. It demonstrates exceptional scalability, processing a 1-million-point scan in approximately 13 minutes, compared to 90 minutes for competing global optimization methods like QPBO. Furthermore, the oriented normals produced by our method lead to superior downstream results in Poisson Surface Reconstruction (PSR).
Discussion and Conclusion
By framing the orientation problem as a dipole interaction, we effectively utilize the Poisson equation as a regularizer. The resulting field naturally converges toward the shape's winding number, providing a robust topological measure of interior vs. exterior. The use of neural confidence weighting allows certain regions of the geometry to guide the orientation of more ambiguous or noisy areas.
While highly effective for watertight manifolds, the method currently faces challenges with "polygon soups" containing complex internal intersecting surfaces. Future work will explore integrating visibility-based cues from scanner positions to further enhance performance on non-closable surfaces.
Summary updated