Research map / 3D vision / Gaussian splatting
EF-3DGS: Event-Aided Free-Trajectory 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.
Integrates event camera streams with video frames into 3D Gaussian Splatting to jointly reconstruct scenes and optimize camera poses from high-speed, free-trajectory videos.
Problem
Scene reconstruction and pose estimation from casually captured free-trajectory videos fails under high-speed motion, where sparse frames and large inter-frame displacements make joint pose-scene optimization ill-posed.
Approach
Uses event streams to supervise continuous brightness between frames via an event generation model, adds contrast maximization on piece-wise warped events with depth-derived optical flow for pose/gradient constraints, and uses photometric bundle adjustment plus a two-stage Fixed-GS schedule to recover color.
Key results
- On Tanks and Temples at 1 FPS, EF-3DGS reaches 23.96 PSNR versus 20.53 for CF-3DGS (about 3.4 dB gain) and 22.41 for Event-3DGS.
- At 1 FPS on Tanks and Temples, ATE drops to 0.3753 from 0.8972 for CF-3DGS (about 58% lower), and RPE_t is 0.0487 vs 0.1057.
- On RealEv-DAVIS FAST sequences, EF-3DGS achieves 21.12 PSNR and ATE-related RPE_t 1.80 vs 19.13 and 2.70 for EvCF-3DGS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add EGM event supervision (ablation) | EF-3DGS · Tanks and Temples 1 FPS | PSNR: 20.53 -> 22.16 | Helped |
| add LEGM motion and gradient constraints (ablation) | EF-3DGS · Tanks and Temples 1 FPS | PSNR: 20.53 -> 21.07 | Helped |
| add PBA cross-view photometric consistency (ablation) | EF-3DGS · Tanks and Temples 1 FPS | PSNR: 20.53 -> 20.96 | Helped |
| add both EGM and LEGM (ablation) | EF-3DGS · Tanks and Temples 1 FPS | ATE: 0.8972 -> 0.3981 | Helped |
| add Fixed-GS to EGM and LEGM (ablation) | EF-3DGS · Tanks and Temples 1 FPS | PSNR: 22.83 -> 23.46 | Helped |
| add Fixed-GS to EGM, LEGM, and PBA (ablation) | EF-3DGS · Tanks and Temples 1 FPS | ATE: 0.3753 -> 0.3753 | No clear effect |
| set CMax r to 1 instead of 3 (ablation) | EF-3DGS · Tanks and Temples 1 FPS | PSNR: 23.96 -> 22.94 | Hurt |
| set CMax r to 5 instead of 3 (ablation) | EF-3DGS · Tanks and Temples 1 FPS | PSNR: 23.96 -> 23.49 | Hurt |
| use reference-view color loss each iteration instead of two-stage Fixed-GS (ablation) | EF-3DGS · Tanks and Temples color correction | Training Time: 2.5h -> 3.8h | Hurt |
| use two-stage Fixed-GS rather than reference-view color loss (ablation) | EF-3DGS · Tanks and Temples color correction | PSNR: 24.11 vs 23.96 | Mixed |
| use event-aided EF-3DGS rather than CF-3DGS | 3DGS reconstruction · RealEv-DAVIS FAST | PSNR: 17.59 -> 21.12 | Helped |
| use event-aided EF-3DGS rather than Event-3DGS | 3DGS reconstruction · Tanks and Temples 1 FPS | PSNR: 22.41 -> 23.96 | Helped |
Surprising findings
- Pose-free event-based EF-3DGS outperforms Event-3DGS even though Event-3DGS uses precomputed COLMAP poses.
- Naively adding the event loss without a color-fixing stage causes color distortion, since dense event supervision dominates sparse RGB supervision.
Limitations
- Event-based pipeline depends on event simulation (V2E) for benchmark data, and real-world event data is limited to one collected dataset.
- Gains depend on the number of subintervals N and warping window r, which are tuned per frame rate; the optimal N may vary across motion characteristics.
- Does not explicitly model motion blur; blur extension is only supplementary with EDI.
Experimental setup
Datasets: Tanks and Temples RealEv-DAVIS
Benchmarks: Tanks and Temples RealEv-DAVIS
Metrics: PSNR SSIM LPIPS ATE RPE_t RPE_r
Baselines: F2-NeRF Nope-NeRF LocalRF CF-3DGS EvDeblurNeRF ENeRF Event-3DGS EvCF-3DGS
Builds on: 3D Gaussian Splatting CF-3DGS LocalRF Contrast Maximization Event Double Integral
Closest papers by meaning
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.93)
- EvGGS: A Collaborative Learning Framework for Event-based Generalizable Gaussian Splatting (ICML 2024 · similarity 0.93)
- GS2E: Gaussian Splatting is an Effective Data Generator for Event Stream Generation (NeurIPS 2025 · similarity 0.91)
- IncEventGS: Pose-Free Gaussian Splatting from a Single Event Camera (CVPR 2025 · similarity 0.90)
- 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)
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.