Research map / 3D vision / Gaussian splatting
Event-3DGS: Event-based 3D Reconstruction Using 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 Event-3DGS, the first event-camera 3D reconstruction framework built on 3D Gaussian Splatting, using a high-pass-filtered photovoltage contrast estimate and a tolerance-based event loss to synthesize novel views from event streams.
Problem
Event-based 3D reconstruction with NeRF is slow in training and inference, has limited scene editability, and is fragile under real-world noise, fast motion, and low light.
Approach
Render log-intensity contrasts from 3DGS and compare them with event-derived photovoltage contrast, which is high-pass filtered to suppress noise, using a tolerance-based (beta) L1-style event loss plus an intensity loss from E2VID initialization, while jointly optimizing the scene and the sensor threshold.
Key results
- On the DeepVoxels synthetic dataset, Event-3DGS reaches average 21.762 PSNR and 0.925 SSIM, versus 19.535 PSNR for E2VID+3DGS.
- On the real Event-Camera dataset, Event-3DGS averages 13.167 PSNR and 0.417 SSIM, versus 6.670 PSNR for E2VID+3DGS and 6.643 PSNR for E2VID.
- Appendix reports about 5 minutes (roughly 500 s at 7999 iterations) of training on synthetic data, versus NeRF-based methods described as slower.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Event-3DGS instead of E2VID+3DGS | 3D reconstruction from event streams · DeepVoxels synthetic (Average) | SSIM: 0.908 -> 0.925 | Helped |
| Use Event-3DGS instead of E2VID+3DGS | 3D reconstruction from event streams · Event-Camera real-world (Average) | PSNR: 6.670 -> 13.167 | Helped |
| Use high-pass filtering rather than pure integration (ablation) | Event-3DGS on real-world event data · Event-Camera real-world (Average) | LPIPS: 0.580 -> 0.245 | Helped |
| Set loss weight α to 0.9 rather than α=1 (ablation) | Event-3DGS · DeepVoxels synthetic | SSIM: 0.945 -> 0.956 | Helped |
| Set contrast tolerance β to 0.5 (ablation) | Event-3DGS · Event-Camera real-world | SSIM: 0.405 -> 0.497 | Helped |
Surprising findings
- On the synthetic benchmark, the pure-integration baseline PI-3DGS reaches higher PSNR on some sequences (e.g., lego 23.853 vs 23.046), and Event-3DGS is only marginally better on average SSIM and PSNR (0.925 vs 0.921 SSIM), indicating limited gain on clean data.
- On the real dataset, Event-3DGS is worse than PI-3DGS on slider_depth in PSNR (12.448 vs 13.509).
Limitations
- Adaptive threshold is learned per scene and does not model temporal variation of the threshold within a scene.
- The 3DGS representation does not support dynamic scenes; 4DGS is suggested as a remedy.
- Relies on E2VID for initial intensity estimation and on a tolerance hyperparameter beta that is set differently for synthetic and real data.
Experimental setup
Datasets: DeepVoxels synthetic dataset Event-Camera dataset
Benchmarks: DeepVoxels synthetic dataset Event-Camera dataset
Metrics: PSNR SSIM LPIPS
Baselines: E2VID E2VID+3DGS PI-3DGS Ev-NeRF
Builds on: 3D Gaussian Splatting E2VID DVS event generation model
Closest papers by meaning
- EvGGS: A Collaborative Learning Framework for Event-based Generalizable Gaussian Splatting (ICML 2024 · similarity 0.93)
- EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering (CVPR 2025 · similarity 0.93)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.93)
- GS2E: Gaussian Splatting is an Effective Data Generator for Event Stream Generation (NeurIPS 2025 · similarity 0.91)
- SpikeGS: Reconstruct 3D Scene Captured by a Fast-Moving Bio-Inspired Camera (AAAI 2025 · similarity 0.89)
- DGS-LRM: Real-Time Deformable 3D Gaussian Reconstruction From Monocular Videos (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.