Research map / 3D vision / NeRF and radiance fields
Exploiting Deblurring Networks for Radiance Fields
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 DeepDeblurRF, which alternates DNN-based (NAFNet) RF-guided deblurring with radiance field construction (Plenoxels or 3D Gaussians) to deblur novel-view synthesis from blurred training views, and releases the BlurRF-Synth dataset.
Problem
Radiance field construction from blurred training views, where linear blur models fail on non-linear outliers (saturation, noise, in-camera processing) and existing ray-based methods are slow to train.
Approach
Combine feed-forward DNN deblurring (NAFNet) with radiance field construction in an iterative loop: the radiance field renders sharper views of each input that guide a RF-guided deblurring network, whose outputs are then used to retrain the radiance field, with the loop repeated N=5 times.
Key results
- On BlurRF-Synth camera motion blur, DeepDeblurRF-G reaches 29.94 PSNR / 0.1059 LPIPS in 0.28 hours, versus 28.33 PSNR for PDRF-10 in 4.26 hours.
- On BlurRF-Synth defocus blur, DeepDeblurRF-G reaches 32.58 PSNR, compared with 30.03 for Deblur-NeRF and 29.90 for BAGS.
- On BlurRF-SB (blur directions identical across views), DeepDeblurRF-P achieves 29.45 PSNR versus 27.07 for DP-NeRF, showing robustness when views lack complementary blur.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeepDeblurRF-P instead of PDRF-10 | Radiance field deblurring on BlurRF-Synth test camera motion blur · BlurRF-Synth test, camera motion blur | PSNR: 28.33 -> 29.81 | Helped |
| Use DeepDeblurRF-G instead of PDRF-10 | Radiance field deblurring on BlurRF-Synth test camera motion blur · BlurRF-Synth test, camera motion blur | PSNR: 28.33 -> 29.94 | Helped |
| Use DeepDeblurRF-P instead of PDRF-10 | Radiance field deblurring on BlurRF-Synth test defocus blur · BlurRF-Synth test, defocus blur | PSNR: 30.03 -> 32.51 | Helped |
| Use DeepDeblurRF-G instead of PDRF-10 | Radiance field deblurring on BlurRF-Synth test defocus blur · BlurRF-Synth test, defocus blur | PSNR: 30.03 -> 32.58 | Helped |
| Use DeepDeblurRF-P instead of PDRF-10 | Radiance field deblurring training on BlurRF-Synth camera motion blur · BlurRF-Synth test, camera motion blur | computation time (Hr.): 4.26 -> 1.14 | Helped |
| Use DeepDeblurRF-G instead of BAGS | Radiance field deblurring training on BlurRF-Synth camera motion blur · BlurRF-Synth test, camera motion blur | computation time (Hr.): 1.25 -> 0.28 | Helped |
| Use DeepDeblurRF-G instead of BAGS | Novel-view synthesis on BlurRF-Real · BlurRF-Real | NIQE: 5.837 -> 5.423 | Helped |
| Use DeepDeblurRF-G instead of BAGS | Novel-view synthesis on BlurRF-Real · BlurRF-Real | ARNIQA: 0.290 -> 0.329 | Helped |
| Use DeepDeblurRF-G instead of DP-NeRF | Novel-view synthesis on BlurRF-SB · BlurRF-SB | PSNR: 27.07 -> 29.59 | Helped |
| Increase iteration count from N=1 to N=6 (ablation) | DeepDeblurRF-G · BlurRF-Synth test, camera motion blur | PSNR: 28.40 -> 29.97 | Helped |
| Increase iteration count from N=1 to N=6 (ablation) | DeepDeblurRF-G · BlurRF-Synth test, defocus blur | PSNR: 30.74 -> 32.75 | Helped |
Surprising findings
- Naive combination of single-image deblurring with radiance field construction gives unsatisfactory results, motivating the RF-guided iterative design.
Limitations
- Struggles to restore images with blur more severe than in the training data.
- RF-guided deblurring uses only the rendered image of each input view for guidance, without additional priors from other views.
Experimental setup
Datasets: BlurRF-Synth Deblur-NeRF benchmark
Benchmarks: BlurRF-Synth BlurRF-Real BlurRF-SB Deblur-NeRF real and synthetic scenes
Metrics: PSNR SSIM LPIPS NIQE ARNIQA computation time
Baselines: Deblur-NeRF BAD-NeRF DP-NeRF ExBluRF PDRF-10 Deblurring-3DGS BAGS
Builds on: NAFNet Plenoxels 3D Gaussian Splatting Deblur-NeRF RSBlur
Closest papers by meaning
- Deblur-NeRF: Neural Radiance Fields From Blurry Images (CVPR 2022 · similarity 0.91)
- DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video (CVPR 2024 · similarity 0.89)
- Efficient Dynamic Scene Deblurring Using Spatially Variant Deconvolution Network With Optical Flow Guided Training (CVPR 2020 · similarity 0.89)
- DynaMoDe-NeRF: Motion-aware Deblurring Neural Radiance Field for Dynamic Scenes (CVPR 2025 · similarity 0.87)
- Bright-NeRF: Brightening Neural Radiance Field with Color Restoration from Low-Light RAW Images (AAAI 2025 · similarity 0.87)
- Mitigating Motion Blur in Neural Radiance Fields with Events and Frames (CVPR 2024 · similarity 0.87)
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.