Research map / 3D vision / Gaussian splatting
EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering
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 EventSplat, a 3D Gaussian Splatting method trained directly from event-accumulated log-difference images with event-to-video guided SfM initialization and cubic spline pose interpolation for novel view synthesis from moving event cameras.
Problem
Novel view synthesis from moving event cameras is slow with event-based NeRFs, and 3DGS normally requires RGB supervision, so event data cannot directly supervise it; sparse events and unknown high-rate poses also hamper reconstruction.
Approach
Supervise 3DGS with accumulated event images by rasterizing two views, taking their log difference (with Bayer remosaicing for color), initialize Gaussians from SfM on event-to-video reconstructed frames, and get dense poses via cubic spline interpolation of sparse tracked poses.
Key results
- On synthetic scenes, mean PSNR 28.14 vs 28.19 for Robust-e-NeRF and mean SSIM 0.953 vs 0.945, with rendering in a few milliseconds versus tens of ms for Robust-e-NeRF.
- On real EDS scenes, mean PSNR 18.86 and SSIM 0.792, versus 16.25 and 0.739 for Robust-e-NeRF.
- Ablation on real scenes: event-to-video initialization raised PSNR from 18.02 to 18.75; adding cubic spline interpolation gave 18.86.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add event-to-video guided initialization instead of random initialization (ablation) | EventSplat · EDS 07_ziggy_and_fuzz_hdr | PSNR: 18.02 -> 18.75 | Helped |
| add event-to-video guided initialization instead of random initialization (ablation) | EventSplat · EDS 07_ziggy_and_fuzz_hdr | SSIM: 0.767 -> 0.788 | Helped |
| add event-to-video guided initialization instead of random initialization (ablation) | EventSplat · EDS 07_ziggy_and_fuzz_hdr | LPIPS: 0.397 -> 0.359 | Helped |
| add cubic spline interpolation to event-to-video guided initialization (ablation) | EventSplat · EDS 07_ziggy_and_fuzz_hdr | PSNR: 18.75 -> 18.86 | Helped |
| add cubic spline interpolation to event-to-video guided initialization (ablation) | EventSplat · EDS 07_ziggy_and_fuzz_hdr | SSIM: 0.788 -> 0.792 | Helped |
| add cubic spline interpolation to event-to-video guided initialization (ablation) | EventSplat · EDS 07_ziggy_and_fuzz_hdr | LPIPS: 0.359 -> 0.363 | Hurt |
Surprising findings
- On synthetic scenes the mean PSNR of EventSplat (28.14) is essentially tied with Robust-e-NeRF (28.19), with Robust-e-NeRF higher on several scenes.
Limitations
- Limited to static scenes; cannot handle dynamic moving objects.
- Reconstructs relative log-intensity, requiring a linear color transform with reference data at inference.
- Narrow field of view of event cameras causes border artifacts; only the EDS dataset provides quantitative real-scene evaluation.
Experimental setup
Datasets: ESIM synthetic scenes EDS TUM-VIE
Benchmarks: ESIM synthetic scenes (7 scenes) EDS real sequences TUM-VIE mocap sequences
Metrics: PSNR SSIM LPIPS rendering time (ms)
Baselines: Robust-e-NeRF E2VID+3DGS
Builds on: 3D Gaussian Splatting E2VID Robust-e-NeRF ESIM
Closest papers by meaning
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- GS2E: Gaussian Splatting is an Effective Data Generator for Event Stream Generation (NeurIPS 2025 · similarity 0.92)
- EvGGS: A Collaborative Learning Framework for Event-based Generalizable Gaussian Splatting (ICML 2024 · similarity 0.90)
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- COLMAP-Free 3D Gaussian Splatting (CVPR 2024 · similarity 0.89)
- Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis (CVPR 2024 · 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.