Research map / 3D vision / Gaussian splatting
GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring
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.
Builds GS-Blur, a synthetic deblurring dataset rendered from 3D Gaussian Splatting reconstructions of MVImgNet objects, with random Bezier camera trajectories, averaging sub-frames, and noise/multi-resolution augmentation, and shows models trained on it generalize better across GoPro, REDS, BSD, and RSBlur.
Problem
Existing deblurring datasets are either synthetic (averaged high-speed frames) with limited blur diversity or real beam-splitter captures that are camera-specific and labor-intensive, so they generalize poorly to real-world blur.
Approach
Reconstruct 3D scenes with 3DGS from multi-view object videos, then synthesize blur by averaging dense sub-frame renderings along randomly generated 6DOF Bezier camera trajectories, plus rigid object motion, sensor noise, and multi-resolution rendering, enabling diverse blur and 1-to-n blur generation.
Key results
- Models trained on GS-Blur achieve best cross-dataset PSNR in most settings, e.g., NAFNet on RSBlur 32.30 dB vs 30.63 dB when trained on BSD.
- PSNR thresholding for 3DGS scene filtering improves NAFNet on GoPro from 31.46 to 31.49 and RSBlur from 32.13 to 32.30 vs unfiltered.
- Ablation shows removing rigid object motion, 1-to-n generation, noise, or multi-resolution each lowers performance on at least some benchmarks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on GS-Blur rather than GoPro, REDS, BSD, or RSBlur | MIMO-UNet, UFormer, and NAFNet · GoPro, REDS, BSD, and RSBlur | PSNR | Helped |
| Add 1-to-n blur generation (ablation) | GS-Blur-trained NAFNet · GoPro, REDS, BSD, and RSBlur | PSNR: 30.69 -> 31.49; 30.40 -> 30.54; 29.69 -> 31.37; 30.18 -> 32.30 | Helped |
| Add rigid-body object motion blur (ablation) | GS-Blur-trained NAFNet · GoPro, REDS, BSD, and RSBlur | PSNR: 31.44 -> 31.49; 30.27 -> 30.54; 31.11 -> 31.37; 32.29 -> 32.30 | Helped |
| Add realistic noise (ablation) | GS-Blur-trained NAFNet · GoPro, BSD, and RSBlur | PSNR: 30.97 -> 31.49; 30.36 -> 31.37; 30.99 -> 32.30 | Helped |
| Add realistic noise (ablation) | GS-Blur-trained NAFNet · REDS | PSNR: 30.94 -> 30.54 | Hurt |
| Add multi-resolution renderings (ablation) | GS-Blur-trained NAFNet · GoPro, REDS, BSD, and RSBlur | PSNR: 31.24 -> 31.49; 30.07 -> 30.54; 31.06 -> 31.37; 32.23 -> 32.30 | Helped |
| Filter scenes using PSNR thresholding (ablation) | GS-Blur dataset trained with NAFNet · GoPro, REDS, BSD, and RSBlur | PSNR: 31.46 -> 31.49; 30.52 -> 30.54; 31.42 -> 31.37; 32.13 -> 32.30 | Mixed |
Surprising findings
- Adding noise lowers performance on REDS because REDS suppresses noise by downscaling its evaluation set.
Limitations
- Static 3D scenes cannot model deforming objects such as walking limbs or rotating wheels.
- Sharp images are rendered from 3DGS, which may differ from real clean camera images.
- Only 26 MVImgNet object classes selected, mostly object-centric scenes.
Experimental setup
Datasets: MVImgNet GoPro REDS BSD RSBlur
Benchmarks: GoPro REDS BSD RSBlur real blurry images
Metrics: PSNR SSIM MUSIQ TOPIQ
Baselines: MIMO-UNet Uformer NAFNet
Builds on: 3D Gaussian Splatting MVImgNet RSBlur noise pipeline
Closest papers by meaning
- BARD-GS: Blur-Aware Reconstruction of Dynamic Scenes via Gaussian Splatting (CVPR 2025 · similarity 0.91)
- DiET-GS: Diffusion Prior and Event Stream-Assisted Motion Deblurring 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- DOF-GS: Adjustable Depth-of-Field 3D Gaussian Splatting for Post-Capture Refocusing, Defocus Rendering and Blur Removal (CVPR 2025 · similarity 0.88)
- Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization (CVPR 2024 · similarity 0.88)
- Blur2Blur: Blur Conversion for Unsupervised Image Deblurring on Unknown Domains (CVPR 2024 · similarity 0.88)
- BlurDM: A Blur Diffusion Model for Image Deblurring (NeurIPS 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.