Research map / 3D vision / Gaussian splatting
DeSplat: Decomposed Gaussian Splatting for Distractor-Free 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.
Decomposes 3D Gaussian splatting into static Gaussians and per-view distractor Gaussians trained purely with photometric loss, enabling distractor-free rendering without pre-trained semantic models.
Problem
Distractors and occluders break multi-view consistency, causing floaters and artifacts when reconstructing real-world scenes with 3D Gaussian splatting; existing fixes rely on pre-trained semantic networks or extra MLPs that add cost.
Approach
Initialize a separate set of distractor Gaussians per training view (placed on a plane in front of the camera) alongside global static Gaussians, alpha-composite the two renderings, and regularize opacities so distractors are explained by view-specific Gaussians while the static scene stays clean; no external networks or semantic features are used.
Key results
- On RobustNeRF, DeSplat achieves 23.40 PSNR on Statue and 36.15 PSNR on Crab (1), best or on par with baselines across scenes.
- On On-the-go, DeSplat performs best or on par in medium and high occlusion scenes (e.g., Patio-High 22.59 PSNR, Spot 26.07 PSNR) but underperforms on Mountain and Fountain low-occlusion outdoor scenes.
- Ablation on Statue shows distractor Gaussians with ADC and regularization raise PSNR from 22.75 (Splatfacto) to 23.40, and DeSplat keeps performance stable across distractor ratios where Splatfacto drops.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add per-view distractor Gaussians without ADC or regularization (ablation) | DeSplat · RobustNeRF — Statue | PSNR: 22.75 -> 21.37 | Hurt |
| Add per-view distractor Gaussians without ADC or regularization (ablation) | DeSplat · RobustNeRF — Android | PSNR: 24.46 -> 22.92 | Hurt |
| Add ADC for distractor Gaussians (ablation) | DeSplat with distractor Gaussians and no regularization · RobustNeRF — Statue | PSNR: 21.37 -> 23.52 | Helped |
| Add ADC for distractor Gaussians (ablation) | DeSplat with distractor Gaussians and no regularization · RobustNeRF — Android | PSNR: 22.92 -> 24.69 | Helped |
| Add opacity regularization (ablation) | DeSplat with distractor ADC · RobustNeRF — Statue | PSNR: 23.52 -> 23.40 | Hurt |
| Add opacity regularization (ablation) | DeSplat with distractor ADC · RobustNeRF — Android | PSNR: 24.69 -> 24.80 | Helped |
Surprising findings
- Without ADC, adding distractor Gaussians lowers PSNR below the Splatfacto baseline on Statue (21.37 vs 22.75).
Limitations
- Distractors static across most views may be mistakenly treated as static scene content.
- Small colour differences between distractors and static objects hinder separation; outdoor skies and lighting changes are hard.
- Struggles on low-occlusion outdoor scenes (Mountain, Fountain), with PSNR penalized by holes.
Experimental setup
Datasets: RobustNeRF On-the-go Photo Tourism
Benchmarks: RobustNeRF On-the-go Photo Tourism
Metrics: PSNR SSIM LPIPS
Baselines: Splatfacto Splatfacto-W SpotlessSplats WildGaussians RobustNeRF NeRF On-the-go NeRF-HuGS SWAG GS-W Wild-GS
Builds on: 3D Gaussian Splatting Splatfacto Adaptive Density Control NeRF-W RobustNeRF
Closest papers by meaning
- FreeSplat: Generalizable 3D Gaussian Splatting Towards Free View Synthesis of Indoor Scenes (NeurIPS 2024 · similarity 0.91)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Swift4D: Adaptive divide-and-conquer Gaussian Splatting for compact and efficient reconstruction of dynamic scene (ICLR 2025 · similarity 0.90)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.90)
- HiSplat: Hierarchical 3D Gaussian Splatting for Generalizable Sparse-View Reconstruction (ICLR 2025 · similarity 0.90)
- Speedy-Splat: Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives (CVPR 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.