Research map / Image generation / Image restoration
Hyper-MD: Mesh Denoising with Customized Parameters Aware of Noise Intensity and Geometric Characteristics
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 Hyper-MD, a hyper-network that customizes the parameters of a facet-normal denoising network per facet from two angles capturing noise intensity and geometric characteristics, using coarse-grained base models plus predicted fine-grained offsets.
Problem
Deep-learning mesh denoising networks ignore the diversity of geometric characteristics and noise distributions across mesh facets, limiting denoising quality.
Approach
A hyper-network takes two angles computed on a filtered mesh (noise intensity via normal gap to filtered normal, geometric characteristics via max normal difference in 2-ring) and outputs parameters for the MD network as a coarse-grained base chosen from 12 category-specific candidates plus fine-grained predicted offsets.
Key results
- On SynData complex meshes, Hyper-MD achieves average Ea 4.50 and Ev 7.85e-4, the best among compared methods (GCN-Denoiser Ea 5.25).
- On simple meshes, Hyper-MD averages Ea 4.32 vs TGV 3.84 (TGV better on Ea) but better Ev (4.32 vs 3.84 is Ea; Ev 2.20e-3 vs TGV 3.01e-3).
- Ablation shows removing the hyper-network (w/o NM) gives the largest degradation, e.g., complex Ea 5.28 vs 4.50.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Hyper-MD instead of TGV | SynData simple meshes · SynData simple meshes | Ea ↓ (average): 3.84 -> 4.32 | Hurt |
| use Hyper-MD instead of BNF | SynData simple meshes · SynData simple meshes | Ev ↓ (×10−3, average): 2.32 -> 2.20 | Helped |
| use Hyper-MD instead of GCN | SynData complex meshes · SynData complex meshes | Ea ↓ (average): 5.25 -> 4.50 | Helped |
| use Hyper-MD instead of GCN | SynData complex meshes · SynData complex meshes | Ev ↓ (×10−4, average): 8.21 -> 7.85 | Helped |
| add the hyper-network NM and customized parameters (ablation) | Hyper-MD · SynData simple meshes | Ea: 4.91 -> 4.32 | Helped |
| add the hyper-network NM and customized parameters (ablation) | Hyper-MD · SynData complex meshes | Ea: 5.28 -> 4.50 | Helped |
| add the coarse-grained base (ablation) | Hyper-MD · SynData simple meshes | Ea: 4.84 -> 4.32 | Helped |
| add the coarse-grained base (ablation) | Hyper-MD · SynData complex meshes | Ea: 4.83 -> 4.50 | Helped |
| add the fine-grained offsets (ablation) | Hyper-MD · SynData simple meshes | Ea: 4.46 -> 4.32 | Helped |
| add αn to the hyper-network input (ablation) | Hyper-MD · SynData complex meshes | Ea: 4.55 -> 4.50 | Helped |
| add αg to the hyper-network input (ablation) | Hyper-MD · SynData simple meshes | Ea: 4.37 -> 4.32 | Helped |
Surprising findings
- Hyper-MD is roughly 2.4x slower than GCN-Denoiser at denoising (920s vs 384s on SynData) due to more iterations.
Limitations
- Many hyperparameters from specialized designs need tuning.
- Training is laborious and time-consuming (about 80 hours).
- Facet-wise iterative denoising incurs significant time overhead.
Experimental setup
Datasets: SynData Kv1Data Kv2Data K-FData
Benchmarks: SynData Kv1Data Kv2Data K-FData
Metrics: Ea (average normal angular difference) Ev (normalized average Hausdorff vertex error)
Baselines: BMF BNF GNF TGV GCN-Denoiser
Builds on: GCN-Denoiser EdgeConv Meta-PU MetaAnchor
Closest papers by meaning
- Local Surface Descriptor for Geometry and Feature Preserved Mesh Denoising (AAAI 2022 · similarity 0.86)
- Topological and Node Noise Filtering on 3D Meshes Using Graph Neural Networks (Student Abstract) (AAAI 2024 · similarity 0.85)
- You Should Learn to Stop Denoising on Point Clouds in Advance (AAAI 2025 · similarity 0.84)
- U-CAN: Unsupervised Point Cloud Denoising with Consistency-Aware Noise2Noise Matching (NeurIPS 2025 · similarity 0.83)
- Fast, Accurate Manifold Denoising by Tunneling Riemannian Optimization (ICML 2025 · similarity 0.82)
- DMV3D: Denoising Multi-view Diffusion Using 3D Large Reconstruction Model (ICLR 2024 · similarity 0.82)
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.