Research map / 3D vision / NeRF and radiance fields
DyBluRF: Dynamic Neural Radiance Fields from Blurry Monocular Video
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 DyBluRF, a dynamic NeRF that renders sharp novel views from blurry monocular video by jointly modeling camera trajectories, object DCT motion trajectories, cross-time rendering, and extreme-value depth/flow constraints.
Problem
Dynamic NeRF methods assume sharp inputs and degrade badly on monocular videos with motion blur caused by camera and object motion.
Approach
Model the exposure-time image formation explicitly: learn camera poses at discretized exposure timestamps and per-point DCT motion trajectories (learnable basis), render across global time with cross-time rendering, and apply extreme-value constraints (min depth, max flow) on depth and flow priors computed from blurry frames.
Key results
- DyBluRF achieves 25.66 PSNR, 0.895 SSIM, 0.079 LPIPS on the blurry dynamic scenes, outperforming NSFF (23.33 PSNR) and RoDynRF (19.28 PSNR).
- Best of the deblur-then-NeRF pipelines ([56]+[21]) reaches 23.79 PSNR, still below DyBluRF.
- Ablation: removing cross-time rendering drops PSNR to 24.54 from 25.66.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DyBluRF to model blur in 3D rather than a dynamic NeRF baseline on blurry inputs | Dynamic NeRF novel-view synthesis · Stereo Blur Dataset (6-scene average) | PSNR: 23.33 -> 25.66 | Helped |
| Use DyBluRF to model blur in 3D rather than a dynamic NeRF baseline on blurry inputs | Dynamic NeRF novel-view synthesis · Stereo Blur Dataset (6-scene average) | SSIM: 0.834 -> 0.895 | Helped |
| Use DyBluRF to model blur in 3D rather than a dynamic NeRF baseline on blurry inputs | Dynamic NeRF novel-view synthesis · Stereo Blur Dataset (6-scene average) | LPIPS: 0.181 -> 0.079 | Helped |
| Add cross-time rendering (ablation) | DyBluRF · Stereo Blur Dataset (6-scene average) | PSNR: 24.54 -> 25.66 | Helped |
| Use extreme value constraints instead of averaging for data-driven priors (ablation) | DyBluRF · Stereo Blur Dataset (6-scene average) | PSNR: 25.42 -> 25.66 | Helped |
| Add the static model (ablation) | DyBluRF · Stereo Blur Dataset (6-scene average) | PSNR: 24.67 -> 25.66 | Helped |
| Add scene flow constraints (ablation) | DyBluRF · Stereo Blur Dataset (6-scene average) | PSNR: 24.97 -> 25.66 | Helped |
| Learn the DCT basis instead of fixing it to cosine values (ablation) | DyBluRF · Stereo Blur Dataset (6-scene average) | PSNR: 25.49 -> 25.66 | Helped |
Surprising findings
- Preprocessing blurry inputs with state-of-the-art 2D deblurring (Restormer, JCD) before dynamic NeRF training does not close the gap, and in some cases RoDynRF gets worse.
Limitations
- Evaluated on only 6 scenes curated from the Stereo Blur Dataset.
- Depends on data-driven priors (RAFT, MiDaS) that are themselves inaccurate on blurry images.
- Training takes about two days per scene on a single RTX A6000.
Experimental setup
Datasets: Stereo Blur Dataset
Benchmarks: Stereo Blur Dataset (6 dynamic scenes subset)
Metrics: PSNR SSIM LPIPS
Baselines: HyperNeRF DVS NSFF RoDynRF Restormer + NSFF Restormer + RoDynRF JCD + NSFF JCD + RoDynRF
Builds on: NeRF BAD-NeRF NSFF RAFT MiDaS
Closest papers by meaning
- Mitigating Motion Blur in Neural Radiance Fields with Events and Frames (CVPR 2024 · similarity 0.90)
- DynaMoDe-NeRF: Motion-aware Deblurring Neural Radiance Field for Dynamic Scenes (CVPR 2025 · similarity 0.90)
- Exploiting Deblurring Networks for Radiance Fields (CVPR 2025 · similarity 0.89)
- BLiRF: Bandlimited Radiance Fields for Dynamic Scene Modeling (AAAI 2024 · similarity 0.88)
- Dynamic Gaussian Splatting from Defocused and Motion-blurred Monocular Videos (NeurIPS 2025 · similarity 0.88)
- Joint Optimization of Neural Radiance Fields and Continuous Camera Motion from a Monocular Video (CVPR 2025 · 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.