Research map / 3D vision / Gaussian splatting
Micro-macro Wavelet-based Gaussian Splatting for 3D Reconstruction from Unconstrained Images
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 MW-GS, a Scaffold-GS-based 3DGS method for unconstrained photo collections that decouples global, refined, and intrinsic appearance features, using micro-macro conical-frustum projection, wavelet-based multi-scale sampling, and a hierarchical residual fusion MLP.
Problem
Novel view synthesis and 3D reconstruction from unconstrained internet photo collections with varying appearance and transient occlusions, where global appearance embeddings fail to capture fine-grained local appearance.
Approach
Decompose appearance into global, refined, and intrinsic features; sample refined features from 2D CNN feature maps via narrow and broad conical frustum projections with learnable offsets, plus DWT-based multi-scale sampling, and fuse them with a hierarchical residual MLP network.
Key results
- Outperforms GS-W by 1.41 dB, 1.40 dB, and 1.16 dB PSNR on Brandenburg Gate, Sacre Coeur, and Trevi Fountain respectively (29.37, 24.64, 24.07 PSNR).
- Ablation: removing Wavelet-based Sampling drops PSNR by 0.69 dB on Trevi Fountain; removing HRFN drops 0.68 dB on Brandenburg Gate.
- Rendering speed is reported as roughly 1.5x faster than existing Gaussian-based methods (e.g., 64.43 FPS vs 54.49 FPS for GS-W on Brandenburg Gate at 800x800).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MW-GS instead of GS-W | MW-GS · Brandenburg Gate | PSNR: 27.96 -> 29.37 | Helped |
| Use MW-GS instead of GS-W | MW-GS · Sacre Coeur | PSNR: 23.24 -> 24.64 | Helped |
| Use MW-GS instead of GS-W | MW-GS · Trevi Fountain | PSNR: 22.91 -> 24.07 | Helped |
| Add Wavelet-based Sampling (ablation) | MW-GS · Brandenburg Gate | PSNR: 28.88 -> 29.37 | Helped |
| Add Wavelet-based Sampling (ablation) | MW-GS · Trevi Fountain | PSNR: 23.38 -> 24.07 | Helped |
| Add broad conical-frustum projection to micro projection (ablation) | MW-GS · Brandenburg Gate | PSNR: 28.49 -> 29.37 | Helped |
| Add Hierarchical Residual Fusion Network (ablation) | MW-GS · Brandenburg Gate | PSNR: 28.69 -> 29.37 | Helped |
| Add visibility map (ablation) | MW-GS · Brandenburg Gate | PSNR: 28.66 -> 29.37 | Helped |
| Set wavelet downsampling steps M to 1 instead of 0 (ablation) | MW-GS · Brandenburg Gate | PSNR: 28.88 -> 29.37 | Helped |
| Set wavelet downsampling steps M to 1 instead of 2 (ablation) | MW-GS · Brandenburg Gate | PSNR: 28.94 -> 29.37 | Helped |
| Increase conical-frustum samples ks from 1 to 2 (ablation) | MW-GS · Brandenburg Gate | PSNR: 29.37 -> 29.02 | Hurt |
| Increase conical-frustum samples ks from 1 to 3 (ablation) | MW-GS · Brandenburg Gate | PSNR: 29.37 -> 29.01 | Hurt |
Surprising findings
- Increasing the number of conical-frustum samples per point (k_s = 2 or 3) did not improve results and degraded performance, which the authors attribute to features converging toward a common mean.
Limitations
- Struggles to recover fine ground textures such as roads or sidewalks, producing blurred regions.
- Transient masks are learned unsupervised and can obscure unusual but valid appearances; more advanced segmentation is slow.
Experimental setup
Datasets: Brandenburg Gate Sacre Coeur Trevi Fountain
Benchmarks: Brandenburg Gate Sacre Coeur Trevi Fountain
Metrics: PSNR SSIM LPIPS FPS
Baselines: NeRF-W Ha-NeRF CR-NeRF WildGaussians GS-W
Builds on: 3D Gaussian Splatting Scaffold-GS GS-W NeRF-W mip-nerf
Closest papers by meaning
- Superpoint Gaussian Splatting for Real-Time High-Fidelity Dynamic Scene Reconstruction (ICML 2024 · similarity 0.90)
- Learning Efficient Fuse-and-Refine for Feed-Forward 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Metropolis-Hastings Sampling for 3D Gaussian Reconstruction (NeurIPS 2025 · similarity 0.89)
- Splatter Image: Ultra-Fast Single-View 3D Reconstruction (CVPR 2024 · similarity 0.89)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 2025 · similarity 0.89)
- VA-GS: Enhancing the Geometric Representation of Gaussian Splatting via View Alignment (NeurIPS 2025 · similarity 0.89)
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.