Research map / 3D vision / Gaussian splatting
GS2E: Gaussian Splatting is an Effective Data Generator for Event Stream Generation
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.
Builds GS2E, a synthetic event dataset of over 1,000 scenes by reconstructing sparse multi-view RGB scenes with 3D Gaussian Splatting, generating adaptive velocity-controlled camera trajectories, and simulating events with DVS-Voltmeter at tuned contrast thresholds.
Problem
Event-based vision and event-based 3D reconstruction lack large-scale, high-quality event datasets with viewpoint diversity and geometric consistency; existing video-driven and graphics-engine simulators suffer from narrow baselines, domain gaps, and fixed contrast thresholds.
Approach
Reconstruct photorealistic static scenes from sparse multi-view images with 3DGS, render dense trajectories with velocity-controlled B-spline reparameterization and novel-view augmentation, then convert the renders into events with a stochastic DVS-Voltmeter simulator using tuned contrast thresholds.
Key results
- 3DGS reconstruction on MVImgNet reaches average PSNR 29.8, SSIM 0.92, LPIPS 0.14.
- Event-fusion Event-3DGS achieves PSNR 24.31 under mild speed vs 21.89 for E-NeRF event-fusion, with performance dropping at strong speed (19.36).
- Deblurring: EFNet reports PSNR 31.26 vs D2Net 29.61 on the GS2E-based evaluation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use cubic B-spline interpolation with velocity control instead of linear interpolation (ablation) | GS2E camera trajectory generation · Virtual camera trajectories | trajectory smoothness | Helped |
| Use linear interpolation instead of cubic B-spline interpolation (ablation) | GS2E camera trajectory generation · Virtual camera trajectories | temporal coherence | Hurt |
| Raise the contrast threshold from c ≤ 0.75 to c ≥ 1 (ablation) | GS2E event simulation · 3DGS-simulated event streams | floater artifacts | Helped |
| Use a low contrast threshold, c ≤ 0.75 (ablation) | GS2E event simulation · 3DGS-simulated event streams | floater artifacts | Hurt |
Surprising findings
- Low contrast thresholds (c <= 0.75) produce floater artifacts from background Gaussians, so the authors chose c=1 despite the paper's text stating a [0.2, 0.5] range, an internal inconsistency.
Limitations
- Assumes static scenes; dynamic object motion is not modeled.
- Inherits 3DGS photometric limits and cannot model extreme over- or under-exposure or HDR illumination.
- Event realism validated only qualitatively against DSEC, with no quantitative sim-to-real metric.
Experimental setup
Datasets: MVImgNet DL3DV DSEC
Benchmarks: GS2E
Metrics: PSNR SSIM LPIPS CLIPIQA MUSIQ RANKIQA
Baselines: E-NeRF Event-3DGS D2Net EFNet E2VID TimeLens++
Builds on: 3D Gaussian Splatting DVS-Voltmeter E2VID COLMAP
Closest papers by meaning
- EvGGS: A Collaborative Learning Framework for Event-based Generalizable Gaussian Splatting (ICML 2024 · similarity 0.92)
- EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering (CVPR 2025 · similarity 0.92)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.91)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.91)
- DiET-GS: Diffusion Prior and Event Stream-Assisted Motion Deblurring 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- IncEventGS: Pose-Free Gaussian Splatting from a Single Event Camera (CVPR 2025 · similarity 0.88)
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.