GS-Blur: A 3D Scene-Based Dataset for Realistic Image Deblurring

TL;DR

By adopting various camera trajectories in reconstructing the authors' GS-Blur, the dataset contains realistic and diverse types of blur, offering a large-scale dataset that generalizes well to real-world blur.

Abstract

To train a deblurring network, an appropriate dataset with paired blurry and sharp images is essential. Existing datasets collect blurry images either synthetically by aggregating consecutive sharp frames or using sophisticated camera systems to capture real blur. However, these methods offer limited diversity in blur types (blur trajectories) or require extensive human effort to reconstruct large-scale datasets, failing to fully reflect real-world blur scenarios. To address this, we propose GS-Blur, a dataset of synthesized realistic blurry images created using a novel approach. To this end, we first reconstruct 3D scenes from multi-view images using 3D Gaussian Splatting (3DGS), then render blurry images by moving the camera view along the randomly generated motion trajectories. By adopting various camera trajectories in reconstructing our GS-Blur, our dataset contains realistic and diverse types of blur, offering a large-scale dataset that generalizes well to real-world blur. Using GS-Blur with various deblurring methods, we demonstrate its ability to generalize effectively compared to previous synthetic or real blur datasets, showing significant improvements in deblurring performance.

Figures (12)

• Figure 1: Examples of the proposed GS-Blur dataset. The left half of the frames displays synthetically generated blur, while the right half exhibits sharp pairs.
• Figure 1: Comparison of existing datasets with our new GS-Blur dataset.
• Figure 2: Motion distribution visualization of synthetic, real, and GS-Blur datasets.
• Figure 3: The overall pipeline for generating blurry and sharp image pairs in our GS-Blur dataset. To this end, we first train a 3D Gaussian Splatting model to reconstruct 3D scenes from multi-view images. Then, from these reconstructed 3D scenes and randomly generated motion trajectory $\mathbf{T}$, we render sharp images $\mathcal{I}(\mathbf{P}_{0.5(\tau_o + \tau_c)})$ from a fixed camera view and blurry images $\mathcal{B}(\mathbf{T})$ from a moving camera view. Specifically, we render $\mathcal{M}$ sharp images along the motion trajectory and then average these sharp frames to synthesize the blurry image.
• Figure 3: Quantitative comparison on real blurry images su2017deep.