Research map / 3D vision / Gaussian splatting
EvGGS: A Collaborative Learning Framework for Event-based Generalizable 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 EvGGS, a feedforward, generalizable event-only 3D Gaussian Splatting framework that jointly trains depth, intensity, and Gaussian regression modules in a cascade, plus a new synthetic event 3D dataset (Ev3DS) and a real event dataset (Ev3D-R).
Problem
Reconstructing 3D scenes from raw event streams is hard because events are sparse and carry no absolute color, and existing event-based neural reconstruction methods require per-scene optimization and cannot generalize to unseen scenes.
Approach
A cascaded three-module pipeline (event-based monocular depth UNet, intensity UNet conditioned on depth features, and a Gaussian parameter regressor) trained end-to-end with a joint loss under differentiable 3DGS rendering, so that depth, intensity and 3D Gaussian quality mutually reinforce each other.
Key results
- Novel view synthesis on Ev3DS: EvGGS achieves PSNR 27.95, SSIM 0.968, LPIPS 0.045, outperforming EventNeRF (24.62 PSNR) and E-ENeRF (23.86 PSNR).
- Intensity reconstruction: jointly trained EvGGS reaches PSNR 29.18 vs 27.91 for EVSNN and 27.78 for E2VID; independent training gives 26.94.
- Depth estimation: jointly trained model achieves Abs.rel 39.4 (x1e-3) vs 46.2 for EReFormer, a stated improvement of at least 14.7%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| jointly train the depth, intensity, and Gaussian regression modules instead of training them independently (ablation) | EvGGS · Intensity reconstruction | PSNR: 26.94 -> 29.18 | Helped |
| jointly train the depth, intensity, and Gaussian regression modules instead of training them independently (ablation) | EvGGS · Depth estimation | Abs.rel: 51.5 -> 39.4 | Helped |
| add cascading feature connections between modules (ablation) | EvGGS · Novel view synthesis | PSNR: 26.51 -> 27.95 | Helped |
| add depth supervision L_D (ablation) | EvGGS · Depth estimation | Abs.rel: 49.2 -> 39.4 | Helped |
| add intensity supervision L_I (ablation) | EvGGS · Novel view synthesis | LPIPS: 0.518 -> 0.045 | Helped |
| fine-tune EvGGS on realistic event data | EvGGS · Ev3D-R | PSNR: 26.77 -> 27.84 | Helped |
| use EvGGS rather than EventNeRF | event-based novel view synthesis · Novel view synthesis | PSNR: 24.62 -> 27.95 | Helped |
Surprising findings
- Independently trained intensity module was darker than ground truth, with joint training fixing contrast, while E2VID and the intensity module performed identically when trained alone.
- Rendering speed is 195 FPS only after precomputing the Gaussian point cloud; recomputing per render is much slower.
Limitations
- Replaces spherical harmonics with a single intensity value, limiting view-dependent effects.
- Cannot effectively reconstruct specular metals.
- Per-render Gaussian recomputation makes real-time rendering dependent on precomputation.
Experimental setup
Datasets: Ev3DS (synthetic, rendered with Blender/VisionBlender, events via V2E) EventNeRF dataset MipNeRF 360
Benchmarks: Ev3DS Ev3D-R EventNeRF dataset MipNeRF 360 (Bicycle, Garden)
Metrics: PSNR SSIM LPIPS AbsRel SqRel RMSE MAE FPS
Baselines: EventNeRF E-ENeRF E2VID+3DGS E2VID FireNet EVSNN ASNet EReFormer E2Dpt GrayNeRF
Builds on: 3D Gaussian Splatting E2VID ENeRF V2E EventNeRF
Closest papers by meaning
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · 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 0.92)
- EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering (CVPR 2025 · similarity 0.90)
- DiET-GS: Diffusion Prior and Event Stream-Assisted Motion Deblurring 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- SpikeGS: Reconstruct 3D Scene Captured by a Fast-Moving Bio-Inspired Camera (AAAI 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.