Research map / 3D vision / Gaussian splatting
HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting
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 HybridGS, a hybrid of image-specific 2D Gaussians (transients) and multi-view 3D Gaussians (statics) with multi-view supervision and multi-stage training, for novel view synthesis in scenes with transient objects.
Problem
Novel view synthesis with 3D Gaussian Splatting degrades when casually captured images contain transient objects that violate multi-view consistency; prior fixes rely on semantic features or robust losses without explicit transient modeling.
Approach
Model transient objects as single-view planar 2D Gaussians per image while keeping a multi-view consistent 3D Gaussian scene for statics, decoupled via a learned 2D-Gaussian transient mask, with multi-view co-visible-region supervision and a warm-up/iterative/joint training schedule.
Key results
- On NeRF On-the-go, HybridGS improves average PSNR by 1.10 dB and SSIM by 5.27% over the previous SOTA (SLS-mlp).
- On RobustNeRF, HybridGS reports the best PSNR/SSIM/LPIPS across Statue, Android, Yoda and Crab scenes, e.g. Yoda PSNR 35.32 vs 33.66 for SLS-mlp.
- Ablation on 2D Gaussian count finds 10k 2D Gaussians gives the best PSNR (25.034) and SSIM (0.847) on the Corner scene.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add iterative training between 2DGS and 3DGS (ablation) | HybridGS · On-the-go low (5%) | PSNR: 21.39 -> 21.42 | Helped |
| Add iterative training between 2DGS and 3DGS (ablation) | HybridGS · On-the-go medium (17%) | PSNR: 23.11 -> 23.50 | Helped |
| Add iterative training between 2DGS and 3DGS (ablation) | HybridGS · On-the-go high (26%) | PSNR: 22.55 -> 23.05 | Helped |
| Add joint training of 2DGS and 3DGS (ablation) | HybridGS · On-the-go low (5%) | PSNR: 21.05 -> 21.42 | Helped |
| Add joint training of 2DGS and 3DGS (ablation) | HybridGS · On-the-go medium (17%) | PSNR: 23.18 -> 23.50 | Helped |
| Add joint training of 2DGS and 3DGS (ablation) | HybridGS · On-the-go high (26%) | PSNR: 23.01 -> 23.05 | Helped |
| Add multi-view supervision (ablation) | HybridGS · On-the-go low (5%) | PSNR: 21.02 -> 21.42 | Helped |
| Add multi-view supervision (ablation) | HybridGS · On-the-go medium (17%) | PSNR: 23.06 -> 23.50 | Helped |
| Add multi-view supervision (ablation) | HybridGS · On-the-go high (26%) | PSNR: 22.71 -> 23.05 | Helped |
| Use 10k rather than 1k 2D Gaussians (ablation) | HybridGS · Corner | PSNR: 24.493 -> 25.034 | Helped |
| Use 10k rather than 5k 2D Gaussians (ablation) | HybridGS · Corner | PSNR: 24.733 -> 25.034 | Helped |
| Increase the number of 2D Gaussians from 10k to 15k (ablation) | HybridGS · Corner | PSNR: 25.034 -> 24.512 | Hurt |
Surprising findings
- The original 3DGS on Yoda scene sometimes falls below NeRF-based methods, indicating transients severely degrade 3DGS.
Limitations
- Does not model illumination variation in unconstrained photo collections; Photo Tourism results look overcast-like.
- Relies on sparse multi-view data and COLMAP initialization for the static 3D Gaussians.
Experimental setup
Datasets: NeRF On-the-go RobustNeRF
Benchmarks: NeRF On-the-go RobustNeRF
Metrics: PSNR SSIM LPIPS
Baselines: RobustNeRF NeRF On-the-go 3DGS SpotLessSplats (SLS-mlp)
Builds on: 3D Gaussian Splatting 2D Gaussian Splatting GaussianImage SpotLessSplats
Closest papers by meaning
- FlexGS: Train Once, Deploy Everywhere with Many-in-One Flexible 3D Gaussian Splatting (CVPR 2025 · similarity 0.92)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.92)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.92)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.92)
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.92)
- Fully Explicit Dynamic Gaussian Splatting (NeurIPS 2024 · similarity 0.92)
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.