Research map / 3D vision / Gaussian splatting
EGGS: Exchangeable 2D/3D Gaussian Splatting for Geometry-Appearance Balanced Novel View Synthesis
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 EGGS, a hybrid 2D/3D Gaussian splatting method that renders both Gaussian types in one CUDA rasterizer, switches Gaussian types during training via effective rank, and applies wavelet-based frequency-decoupled supervision with gradient projection to balance appearance and geometry.
Problem
3DGS gives high appearance fidelity but multi-view inconsistent geometry, while 2DGS gives accurate geometry but blurred appearance; prior work uses only one Gaussian type per scene, so the complementary strengths are not jointly exploited.
Approach
A unified hybrid rasterizer renders 3D Gaussians by affine projection and 2D Gaussians by ray-splat intersection in one alpha-blending pass; Gaussians switch type during training based on effective rank, and low/high-frequency (DWT) losses are supervised asymmetrically per Gaussian type with conflict-driven gradient projection.
Key results
- On Mip-NeRF360, EGGS reaches 27.96 PSNR / 0.851 SSIM / 0.192 LPIPS, the best among compared methods (3DGS 27.43 PSNR).
- On DTU, EGGS achieves the best mean PSNR of 33.65 and mean Chamfer Distance of 0.91, behind 2DGS (0.80) and GOF (0.74) on CD.
- On LLFF, EGGS reaches 27.34 PSNR and 0.083 LPIPS versus 26.12 and 0.102 for 3DGS-based FreGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EGGS instead of the strongest listed baseline | Novel view synthesis · Mip-NeRF360 | PSNR: 27.92 -> 27.96 | Helped |
| Use EGGS instead of the strongest listed baseline | Novel view synthesis · LLFF | PSNR: 26.53 -> 27.34 | Helped |
| Use EGGS instead of the strongest listed baseline | Novel view synthesis · Tanks&Temples | PSNR: 23.96 -> 24.41 | Helped |
| Use EGGS instead of GOF | DTU reconstruction · DTU | mCD: 0.74 -> 0.91 | Hurt |
| Use EGGS instead of 3DGS | DTU reconstruction · DTU | PSNR: 32.82 -> 33.65 | Helped |
| remove Frequency-Decoupled Optimization (ablation) | EGGS · LLFF | PSNR: 26.58 -> 27.34 | Hurt |
| apply DWT frequency losses directly rather than using full FDO (ablation) | EGGS · LLFF | PSNR: 26.82 -> 27.34 | Hurt |
| mask conflicting frequency gradients rather than use gradient projection (ablation) | EGGS · LLFF | PSNR: 27.07 -> 27.34 | Hurt |
| set erank threshold to 1.9 rather than 2.05 (ablation) | EGGS · LLFF | PSNR: 26.15 -> 27.24 | Hurt |
| set erank threshold to 1.95 rather than 2.05 (ablation) | EGGS · LLFF | PSNR: 26.23 -> 27.24 | Hurt |
| set erank threshold to 2 rather than 2.05 (ablation) | EGGS · LLFF | PSNR: 26.49 -> 27.24 | Hurt |
| set erank threshold to 2.1 rather than 2.05 (ablation) | EGGS · LLFF | PSNR: 27.24 -> 26.37 | Hurt |
Surprising findings
- Even when all Gaussians are initialized as 2D, about 47.8% become 3D by the end of training, showing types change freely during optimization.
- About 20% of Gaussians still show conflicting low/high-frequency gradients at convergence, despite 45% in early training.
Limitations
- Erank is a heuristic that is not monotonic for all scale configurations and has no theoretical guarantee.
- Rendering FPS is lower than 3DGS and GaussianPro on all reported datasets.
- Geometry comparison is weaker than 2DGS/GOF in Chamfer Distance on DTU.
Experimental setup
Datasets: Mip-NeRF360 LLFF Tanks&Temples DTU OOD-NVS
Benchmarks: Mip-NeRF360 LLFF Tanks&Temples DTU OOD-NVS
Metrics: PSNR SSIM LPIPS Chamfer Distance F1 FPS
Baselines: 3DGS 2DGS GaussianPro GOF FreGS TextureGS SUGAR
Builds on: 3DGS 2DGS FreGS ScaffoldGS
Closest papers by meaning
- IBGS: Image-Based Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Training (NeurIPS 2024 · similarity 0.90)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 2024 · similarity 0.90)
- Anti-Aliased 2D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Decoupling Appearance Variations with 3D Consistent Features in Gaussian Splatting (AAAI 2025 · similarity 0.90)
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.