Research map / 3D vision / Gaussian splatting
IncEventGS: Pose-Free Gaussian Splatting from a Single Event Camera
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 IncEventGS, an incremental pose-free 3D Gaussian Splatting reconstruction system driven solely by a single event camera's stream, using tracking-and-mapping with a continuous se(3) trajectory model and monocular depth bootstrapping.
Problem
Dense 3D scene reconstruction and novel view synthesis from a single event camera without ground-truth camera poses, where prior event-based NeRF methods require known poses.
Approach
Incremental event-only 3D-GS SLAM: each event chunk gets a continuous se(3) trajectory (start/end poses interpolated), tracking estimates the new chunk's motion against the existing map, and a sliding-window bundle adjustment jointly refines Gaussians and trajectories by matching measured log-brightness changes; initialization uses random Gaussians and a monocular depth model (Marigold) to re-init
Key results
- On Replica NVS, IncEventGS reaches average PSNR around 21-25 dB per scene (e.g., 24.31 on room0), outperforming E-NeRF, EventNeRF, Robust e-NeRF and E2VID+COLMAP+3DGS despite those using ground-truth poses.
- Pose ATE on Replica is about 0.045-0.115 cm, versus DEVO 0.138-0.289 cm and E2VID+COLMAP up to 105 cm.
- Depth initialization reduces ATE from 1.534 cm to 0.046 cm on Replica office0-type ablation and improves PSNR from 17.80 to 21.74.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add monocular-depth initialization during bootstrapping (ablation) | IncEventGS · Replica office2 | ATE (cm): 1.534 -> 0.046 | Helped |
| add monocular-depth initialization during bootstrapping (ablation) | IncEventGS · Replica office2 | PSNR: 17.80 -> 21.74 | Helped |
| use event slice window 1k-10k instead of 400k-500k (ablation) | IncEventGS · Replica office2 | ATE (cm): 0.046 -> 0.167 | Hurt |
| use event slice window 10k-50k instead of 400k-500k (ablation) | IncEventGS · Replica office2 | ATE (cm): 0.046 -> 0.079 | Hurt |
| use event slice window 80k-200k instead of 400k-500k (ablation) | IncEventGS · Replica office2 | ATE (cm): 0.046 -> 0.079 | Hurt |
| use event slice window 500k-600k instead of 400k-500k (ablation) | IncEventGS · Replica office2 | ATE (cm): 0.046 -> 0.050 | Hurt |
| use event slice window 600k-700k instead of 400k-500k (ablation) | IncEventGS · Replica office2 | ATE (cm): 0.046 -> 0.214 | Hurt |
Surprising findings
- Depth-based re-initialization from a monocular depth model dramatically improves pose accuracy (ATE 1.534 to 0.046 cm), indicating random initialization leaves geometry too poor for tracking.
Limitations
- Relies on random initialization and a pretrained monocular depth model for bootstrapping, so the 3D structure from short event baselines is initially poor.
- Brightness-only NVS metrics are lower than typical RGB results due to lack of absolute brightness supervision.
- ATE on the real TUM-VIE data (0.1-0.3 cm range) is evaluated on a limited set of sequences; real-data NVS lacks paired ground truth.
Experimental setup
Datasets: Replica TUM-VIE
Benchmarks: Replica TUM-VIE
Metrics: PSNR SSIM LPIPS ATE
Baselines: E-NeRF EventNeRF Robust e-NeRF E2VID+COLMAP+3DGS DEVO ESVO2
Builds on: 3D Gaussian Splatting E2VID Marigold NICE-SLAM EventNeRF
Closest papers by meaning
- EF-3DGS: Event-Aided Free-Trajectory 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Gaussian Splatting SLAM (CVPR 2024 · similarity 0.90)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- EventSplat: 3D Gaussian Splatting from Moving Event Cameras for Real-time Rendering (CVPR 2025 · similarity 0.88)
- GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting (CVPR 2024 · similarity 0.88)
- DG-SLAM: Robust Dynamic Gaussian Splatting SLAM with Hybrid Pose Optimization (NeurIPS 2024 · 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.