Research map / 3D vision / NeRF and radiance fields
Mitigating Motion Blur in Neural Radiance Fields with Events and Frames
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.
Ev-DeblurNeRF fuses motion-blurred color frames with events for sharp NeRF reconstruction, using an event double-integral prior, event supervision, and a learnable event camera response function.
Problem
NeRFs trained on motion-blurred images cannot render sharp novel views, and prior event-frame fusion NeRFs struggle to recover accurate colors or restrict supervision to exposure times and fixed poses.
Approach
Combine a Deblur-NeRF-style blur formation model with event-based single-integral supervision applied across the trajectory, an event double-integral (EDI) color prior at mid-exposure, and a learnable event camera response function (eCRF) that adapts to real sensor non-idealities, plus TensoRF features for faster training.
Key results
- On synthetic Ev-DeblurBlender, average PSNR 31.02 (Ev-DeblurNeRF--) and 30.42 (Ev-DeblurNeRF) vs 26.29 for best image-only PDRF and 23.00 for E-NeRF.
- On real Ev-DeblurCDAVIS, average PSNR 32.96 vs 30.48 for E2NeRF (+2.48dB).
- Training is 6.9x faster than E2NeRF (about 3.5 h vs 24 h on an A100).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Ev-DeblurNeRF instead of EFNet* + NeRF | Ev-DeblurNeRF · Ev-DeblurBlender average | PSNR: 28.99 -> 30.42 | Helped |
| Use Ev-DeblurNeRF instead of E2NeRF | Ev-DeblurNeRF · Ev-DeblurCDAVIS average | PSNR: 30.48 -> 32.96 | Helped |
| Add event supervision Lev (ablation) | Image-only model · Ev-DeblurCDAVIS Figures | PSNR: 27.55 -> 28.24 | Helped |
| Add color-channel event supervision Lev-color (ablation) | Model with Lev · Ev-DeblurCDAVIS Figures | PSNR: 28.24 -> 29.28 | Helped |
| Add EDI supervision (ablation) | Event-supervised model · Ev-DeblurCDAVIS Figures | PSNR: +3.15dB | Helped |
| Add learned event CRF (ablation) | Event-supervised model · Ev-DeblurCDAVIS Figures | PSNR: +1.49dB | Helped |
| Add polarity input to event CRF (ablation) | Model with event CRF · Ev-DeblurCDAVIS Figures | PSNR: +0.74dB | Helped |
| Add explicit feature volumes Vc,f (ablation) | Full Ev-DeblurNeRF · Ev-DeblurCDAVIS Figures | PSNR: 33.03 -> 33.17 | Helped |
Surprising findings
- Adding the learnable event camera response function did not improve performance on simulated data, where the ideal event model holds.
- Ev-DeblurNeRF without eCRF and EDI supervision slightly outperforms the full model on the synthetic benchmark.
Limitations
- Requires events and frames from the same sensor and pixel-aligned (e.g., ColorDAVIS), not applicable to all event cameras.
- Event camera poses are obtained by interpolating known frame poses, which may degrade performance if poses are sparse or inaccurate.
Experimental setup
Datasets: Ev-DeblurBlender Ev-DeblurCDAVIS
Benchmarks: Ev-DeblurBlender Ev-DeblurCDAVIS
Metrics: PSNR LPIPS SSIM
Baselines: Deblur-NeRF BAD-NeRF PDRF DP-NeRF MPRNet + NeRF PVDNet + NeRF EFNet + NeRF E-NeRF E2NeRF EDI + NeRF
Builds on: Deblur-NeRF DP-NeRF E-NeRF EDI TensoRF PDRF
Closest papers by meaning
- DynaMoDe-NeRF: Motion-aware Deblurring Neural Radiance Field for Dynamic Scenes (CVPR 2025 · similarity 0.91)
- AE-NeRF: Augmenting Event-Based Neural Radiance Fields for Non-ideal Conditions and Larger Scenes (AAAI 2025 · similarity 0.90)
- DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video (CVPR 2024 · similarity 0.90)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.89)
- LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes (NeurIPS 2024 · similarity 0.89)
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (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.