Research map / 3D vision / Gaussian splatting
EVolSplat: Efficient Volume-based Gaussian Splatting for Urban 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 EVolSplat, a feed-forward 3D Gaussian Splatting model for urban scenes that predicts Gaussians in a global sparse 3D-CNN volume initialized from monocular depth, with IBR-based color and a hemisphere background model.
Problem
Feed-forward novel view synthesis of driving scenes suffers from multi-view inconsistency, duplicated Gaussians from pixel-aligned prediction, and poor depth from feature matching under small parallax; per-scene 3DGS/NeRF optimization is slow.
Approach
Instead of predicting per-pixel Gaussians, initialize a global point cloud from metric monocular depth, process it with a sparse 3D U-Net to predict Gaussian geometry in a unified volume, color Gaussians via occlusion-aware image-based rendering from 2D reference views, and model sky/distant regions with a generalizable hemisphere of Gaussians.
Key results
- On KITTI-360 (50% drop), EVolSplat reaches 23.26 dB PSNR and 0.797 SSIM, outperforming MuRF (22.77 dB) and EDUS (22.13 dB) among feed-forward baselines.
- Zero-shot on Waymo, EVolSplat achieves 23.43 dB PSNR, above MuRF (23.33) and EDUS (23.18), with 83.81 FPS and 10.41 GB inference memory at full KITTI-360 resolution.
- Ablation on KITTI-360 shows removing IBR drops PSNR from 22.83 to 21.06 dB, and removing the occlusion check drops it to 22.54 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add IBR-based Gaussian color prediction (ablation) | EVolSplat · KITTI-360 | PSNR: 21.06 -> 22.83 | Helped |
| add Gaussian position-offset refinement (ablation) | EVolSplat · KITTI-360 | PSNR: 22.76 -> 22.83 | Helped |
| add visibility maps to color aggregation (ablation) | EVolSplat · KITTI-360 | PSNR: 22.54 -> 22.83 | Helped |
| increase IBR projection window size from 1 to 3 (ablation) | EVolSplat · KITTI-360 | PSNR: 22.54 -> 22.83 | Helped |
| increase IBR projection window size from 3 to 5 (ablation) | EVolSplat · KITTI-360 | LPIPS: 0.190 -> 0.201 | Hurt |
Surprising findings
- Fine-tuning from feed-forward initialization converges in under 1000 steps and reduces Gaussian count, while the paper reports comparable PSNR to 3DGS after convergence.
Limitations
- Handling dynamic agents such as other vehicles is not addressed.
- Relies on off-the-shelf monocular metric depth estimators whose misalignment causes blurry colors without IBR.
- Foreground volume is a fixed axis-aligned box whose range must be adjusted per dataset (e.g., for Waymo).
Experimental setup
Datasets: KITTI-360 Waymo Open Dataset
Benchmarks: KITTI-360 Waymo Open Dataset
Metrics: PSNR SSIM LPIPS FPS memory (GB)
Baselines: MVSNeRF MuRF EDUS MVSplat PixelSplat Nerfacto 3DGS
Builds on: 3D Gaussian Splatting MVSplat EDUS StreetGaussian
Closest papers by meaning
- CityGaussianV2: Efficient and Geometrically Accurate Reconstruction for Large-Scale Scenes (ICLR 2025 · similarity 0.91)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.90)
- Horizon-GS: Unified 3D Gaussian Splatting for Large-Scale Aerial-to-Ground Scenes (CVPR 2025 · similarity 0.90)
- FreeSplat: Generalizable 3D Gaussian Splatting Towards Free View Synthesis of Indoor Scenes (NeurIPS 2024 · similarity 0.90)
- Generative Gaussian Splatting for Unbounded 3D City Generation (CVPR 2025 · similarity 0.90)
- GaussianSpa: An "Optimizing-Sparsifying" Simplification Framework for Compact and High-Quality 3D Gaussian Splatting (CVPR 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.