Research map / 3D vision / Gaussian splatting
SpikeGS: Reconstruct 3D Scene Captured by a Fast-Moving Bio-Inspired 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.
Introduces SpikeGS, a 3D Gaussian Splatting framework that reconstructs scenes from Bayer-pattern spike streams captured by a fast-moving color spike camera, using accumulation rasterization and interval supervision, plus a new real and synthetic spike dataset.
Problem
3D reconstruction from fast-moving cameras suffers from motion blur and 3DGS needs sharp images; existing spike-camera NeRF methods are slow and cannot handle Bayer-pattern color spike streams.
Approach
Integrates Bayer-pattern spike streams into the 3DGS pipeline by embedding a time-accumulation process into the rasterizer (accumulation rasterization) and using spike inter-spike intervals for initialization and geometry supervision (interval loss), with the accumulation loss weight ramped up during training.
Key results
- On synthetic data, SpikeGS reaches 33.387 average PSNR and 0.908 SSIM, about +3.47 dB PSNR over the best baseline (TFI+3DGS at 29.916).
- Depth estimation on synthetic data: PSNR 27.360 vs 22.72 for the best baseline (TFI+3DGS).
- On real-world high-speed spike data, SpikeGS achieves average NIQE 9.0912 and IL-NIQE 69.4726, better than the spike baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add interval loss (ablation) | SpikeGS · Synthetic scenes, average | PSNR: 34.12 -> 34.65 | Helped |
| add accumulation loss (ablation) | SpikeGS · Synthetic scenes, average | PSNR: 31.18 -> 34.65 | Helped |
| increase camera poses in accumulation from 3 to 9 (ablation) | SpikeGS · Synthetic scenes, average | PSNR: 33.04 -> 34.65 | Helped |
| increase camera poses in accumulation from 5 to 9 (ablation) | SpikeGS · Synthetic scenes, average | PSNR: 34.14 -> 34.65 | Helped |
| use SpikeGS instead of TFI+3DGS | synthetic datasets · Synthetic datasets, average | PSNR: 29.916 -> 33.387 | Helped |
| use SpikeGS instead of S2I+3DGS | synthetic datasets · Synthetic datasets, average | SSIM: 0.854 -> 0.908 | Helped |
| use SpikeGS instead of TFI+3DGS | synthetic datasets · Synthetic datasets, average | LPIPS: 0.144 -> 0.120 | Helped |
| use SpikeGS instead of Spike-NeRF | real-world datasets · Real-world datasets, average | NIQE: 9.9030 -> 9.0912 | Helped |
| use SpikeGS instead of S2I+3DGS | real-world datasets · Real-world datasets, average | IL-NIQE: 73.9287 -> 69.4726 | Helped |
Limitations
- Real-world evaluation uses only no-reference metrics because no ground truth is available.
- Spike interval supervision suffers from time-domain aliasing and noise sensitivity, so its weight is reduced during training.
- Accumulation over keyframes is an approximation (n<10 keyframes), with error bounded by 1/N.
Experimental setup
Datasets: NeRF synthetic scenes Deblur-NeRF synthetic scenes SpikeGS dataset (real)
Benchmarks: SpikeGS synthetic dataset SpikeGS real-world dataset
Metrics: PSNR SSIM LPIPS NIQE IL-NIQE
Baselines: 3DGS Deblur-3DGS TFI+3DGS TFP+3DGS S2I+3DGS SpikeNeRF Spike-NeRF
Builds on: 3D Gaussian Splatting TFI TFP SCSim
Closest papers by meaning
- USP-Gaussian: Unifying Spike-based Image Reconstruction, Pose Correction and Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Event-3DGS: Event-based 3D Reconstruction Using 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.89)
- WildGaussians: 3D Gaussian Splatting In the Wild (NeurIPS 2024 · similarity 0.89)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.89)
- FlashGS: Efficient 3D Gaussian Splatting for Large-scale and High-resolution Rendering (CVPR 2025 · similarity 0.89)
- GaussianPro: 3D Gaussian Splatting with Progressive Propagation (ICML 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.