Research map / Point clouds / Shape and surface completion
3DMambaIPF: A State Space Model for Iterative Point Cloud Filtering via Differentiable Rendering
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 3DMambaIPF, an iterative Mamba-based encoder-decoder for point cloud filtering with a differentiable point-rendering loss and adaptive ground truth, evaluated on PU-Net and Stanford models with up to 500K points.
Problem
Learning-based point cloud filtering degrades on highly dense, large-scale point clouds and high noise, and produces unrealistic boundaries and outliers.
Approach
Combines Mamba (selective SSM) blocks in an iterative patch-based denoising backbone with a differentiable point-rendering loss on projected views, plus an adaptive ground truth whose noise decreases across iterations.
Key results
- On PU-Net 50K points with 2.5% noise, 3DMambaIPF achieves CD 9.28 and P2M 5.31 (x1e-5), vs IterativePFN 10.15 and 5.88.
- On PU-Net 10K with 2.5% noise, CD 32.62 and P2M 9.92 vs IterativePFN 33.53 and 10.46.
- On Stanford Armadillo (173K points), 3DMambaIPF CD 17.14 is slightly worse than IterativePFN 17.04; Pointfilter has best CD on Armadillo and Dragon.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add differentiable rendering loss (ablation) | 3DMambaIPF · PU-Net 50K, 2.5% noise | CD: 10.15 -> 9.28 | Helped |
| increase rendered views from 8 to 32 (ablation) | 3DMambaIPF · PU-Net 50K, 2.5% noise | CD: 10.39 -> 9.28 | Helped |
| use 6 Mamba layers rather than 1 (ablation) | 3DMambaIPF · PU-Net 50K, 2.5% noise | CD: 13.23 -> 9.28 | Helped |
| increase Mamba layers from 6 to 9 (ablation) | 3DMambaIPF · PU-Net 50K, 2.5% noise | CD: 9.28 -> 9.89 | Hurt |
| replace Mamba backbone with Transformers (ablation) | 3DMambaIPF · PU-Net 50K, 2.5% noise | CD: 9.42 -> 14.02 | Hurt |
| replace Mamba backbone with MLPs (ablation) | 3DMambaIPF · PU-Net 50K, 2.5% noise | CD: 9.42 -> 30.17 | Hurt |
| increase patch size from 1000 to 2000 (ablation) | 3DMambaIPF · PU-Net 50K, 2.5% noise | CD: 9.28 -> 8.98 | Helped |
| use 3DMambaIPF instead of IterativePFN | PU-Net 50K, 1% noise · PU-Net 50K, 1% noise | CD: 6.05 -> 5.89 | Helped |
| use 3DMambaIPF instead of IterativePFN | PU-Net 50K, 2% noise · PU-Net 50K, 2% noise | CD: 8.03 -> 7.55 | Helped |
| use 3DMambaIPF instead of IterativePFN | PU-Net 50K, 2.5% noise · PU-Net 50K, 2.5% noise | CD: 10.15 -> 9.28 | Helped |
| use 3DMambaIPF instead of Pointfilter | Armadillo, 173K points · Armadillo, 173K points | CD: 6.41 -> 17.14 | Hurt |
| use 3DMambaIPF instead of IterativePFN | Dragon, 438K points · Dragon, 438K points | CD: 20.43 -> 19.79 | Helped |
Surprising findings
- On the Stanford Armadillo and Asian-Dragon/Thai-Statue large-scale tests, Pointfilter gets the best quantitative CD/P2M despite poor geometric detail.
- The Mamba backbone with 0.81M parameters outperforms a Transformer variant (158M parameters) in CD and P2M, at lower GFLOPs.
Limitations
- May cause reverse-amplification artifacts on smooth clean surfaces, as the authors acknowledge.
- Does not beat IterativePFN/Pointfilter on all large-scale Stanford metrics (e.g., Armadillo CD slightly worse than IterativePFN).
- Real-world evaluation is only qualitative since Paris-rue-Madame has no GT.
Experimental setup
Datasets: PU-Net Stanford 3D Scanning Repository Paris-rue-Madame
Benchmarks: PU-Net Stanford 3D Scanning Repository Paris-rue-Madame
Metrics: Chamfer Distance (CD) Point-to-Mesh (P2M) distance
Baselines: IterativePFN Pointfilter PD-Flow ScoreDenoise PCN GPDNet DMRDenoise
Builds on: Mamba IterativePFN S4
Closest papers by meaning
- StraightPCF: Straight Point Cloud Filtering (CVPR 2024 · similarity 0.89)
- TIGER: Time-Varying Denoising Model for 3D Point Cloud Generation with Diffusion Process (CVPR 2024 · similarity 0.87)
- Point2Real: Bridging the Gap between Point Cloud and Realistic Image for Open-World 3D Recognition (AAAI 2024 · similarity 0.87)
- SPAR3D: Stable Point-Aware Reconstruction of 3D Objects from Single Images (CVPR 2025 · similarity 0.87)
- Digging into Intrinsic Contextual Information for High-fidelity 3D Point Cloud Completion (AAAI 2025 · similarity 0.87)
- SAMBLE: Shape-Specific Point Cloud Sampling for an Optimal Trade-Off Between Local Detail and Global Uniformity (CVPR 2025 · similarity 0.87)
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.