Research map / Graph learning / GNN architectures
Topological and Node Noise Filtering on 3D Meshes Using Graph Neural Networks (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a GNN operating on DPSR indicator-function grids (128^3) to jointly filter node noise and topological noise (vertex removal) in 3D meshes, with marching cubes for reconstruction.
Problem
Node (positional) noise and topological noise (missing vertices/faces/holes) in 3D meshes are usually handled separately, and GNNs are typically applied only to node noise; treating a mesh as a graph is challenging.
Approach
Convert a mesh to an indicator function grid (via DPSR), then use GNN layers (GATv2-based) on the grid graph to update grid values, followed by marching cubes, so one model handles both node and topological noise.
Key results
- On the synthetic node-noise dataset, the proposed model's VCD/NMCD/AAD/AVD values are reported in Table 1 as lower quality than the best deterministic filters, but it is the only method able to filter topological noise.
- The ablation found two GATv2 layers sufficient and DPSR smoothing coefficient 3.0 best; higher values hurt quality.
- Inference runs on a CPU-only laptop (Intel i7-11370H, 16 GB RAM), reflecting low computational requirements.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use GATv2 layers instead of other graph network layers (ablation) | Proposed mesh denoising model · Mesh denoising | Filtering quality | Helped |
| Use two GATv2 layers (ablation) | Proposed mesh denoising model · Mesh denoising | Filtering quality | No clear effect |
| Increase the DPSR smoothing coefficient above 3.0 (ablation) | Proposed mesh denoising model · Mesh denoising | Filtering quality | Hurt |
| Use the proposed 2GATv2 model instead of established node-noise filters | 2GATv2 · Synthetic dataset (Wang, Liu, and Tong 2016) | VCD: 0,11 vs 11,06 | Hurt |
Surprising findings
- The GATv2 baseline row in Table 1 has far worse values (VCD 11.06, NMCD 16.63) than all other methods, suggesting GNN-only baselines perform poorly on this task.
Limitations
- Only works on smooth, watertight meshes; sharp CAD geometry is poorly represented (Gibbs effect).
- Node-noise quality is lower than several deterministic baselines.
- Evaluated on limited data (synthetic dataset from prior work).
Experimental setup
Datasets: Synthetic dataset from Wang, Liu, and Tong 2016
Benchmarks: Synthetic dataset from Wang, Liu, and Tong 2016
Metrics: VCD NMCD AAD AVD Chamfer distance MSE
Baselines: GATv2 GeoBiGNN CascadedReg GuidedMesh BilaterNorm Fast&Effective
Builds on: Differentiable Poisson Surface Reconstruction (DPSR) GATv2 marching cubes
Closest papers by meaning
- Hyper-MD: Mesh Denoising with Customized Parameters Aware of Noise Intensity and Geometric Characteristics (CVPR 2024 · similarity 0.85)
- Noise-Injected Spiking Graph Convolution for Energy-Efficient 3D Point Cloud Denoising (AAAI 2025 · similarity 0.83)
- PoNQ: a Neural QEM-based Mesh Representation (CVPR 2024 · similarity 0.83)
- Geometric Generative Modeling with Noise-Conditioned Graph Networks (ICML 2025 · similarity 0.83)
- MeshMask: Physics-Based Simulations with Masked Graph Neural Networks (ICLR 2025 · similarity 0.83)
- Joint Graph Rewiring and Feature Denoising via Spectral Resonance (ICLR 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.