Introduction
Creating high-fidelity 3D digital models from real-world scans is often hindered by two types of noise: node noise, which manifests as positional jitter on the surface, and topological noise, which involves missing faces or structural holes. Traditionally, these issues are treated as separate problems—smoothing for jitter and hole-filling for reconstruction. This research proposes a unified approach that addresses both noise types simultaneously using a Graph Neural Network (GNN) framework.
Methodology
The proposed pipeline transforms the irregular 3D mesh into a structured format suitable for deep learning through a three-stage process:
- Grid Conversion ($p_{in}$): The noisy mesh is converted into a $128^3$ indicator function grid. This voxel-based representation indicates whether a point is inside or outside the object, effectively turning the mesh into a consistent graph structure.
- GNN Processing ($\mu_\theta$): The grid is processed by a Graph Neural Network featuring three GATv2 (Graph Attention Network) layers and a Multi-Layer Perceptron (MLP). The attention mechanism allows the network to weigh neighboring voxels differently, enabling it to smooth surfaces and predict missing geometry (hole-filling).
- Mesh Reconstruction ($p_{out}$): The Marching Cubes algorithm extracts a clean, watertight mesh from the processed indicator function.
Experimental Results
The model was evaluated using synthetic datasets and simulated topological noise (e.g., removing 50% of vertices from the "Bunny Hi" model).
Joint Filtering Performance
Unlike standard filters such as BilaterNorm or GuidedMesh, which typically fail when faces are entirely missing, the proposed method successfully reconstructs mesh topology while smoothing the surface. The use of an indicator function allows the model to implicitly "heal" gaps by learning the underlying boundary of the object.
Quantitative Trade-offs
When compared to specialized node-denoising tools, the proposed 2GATv2 model exhibits higher error metrics in scenarios involving only nodal noise. As shown in the table below, metrics like Vertices Chamfer Distance (VCD) are significantly higher for the proposed method compared to specialized algorithms like GeoBiGNN. However, the authors emphasize that this is a necessary trade-off: while specialized tools excel at one task, this model is uniquely capable of handling both node and topological noise in a single pass.
Limitations and Future Directions
While effective for watertight models, the current approach faces challenges with open surfaces and sharp CAD-style geometry due to the "Gibbs effect" inherent in grid-based representations. Additionally, the $128^3$ resolution limits fine detail. Future work aims to utilize octrees for more efficient spatial partitioning, allowing for higher resolutions and the handling of textures and distance functions.
Summary updated