GS-RealBlur beats synthetic blur with flexible real-world data and BPR alignment

Deblurring neural networks have long suffered from a data dilemma: synthetic blur lacks realism, while capturing real blur-sharp pairs requires complex multi-camera rigs. GS-RealBlur, introduced by Mingyang Chen and colleagues, breaks this tradeoff with a simple yet effective setup. They use a handheld camera to capture natural blurry images (e.g., from camera shake or motion) and a gimbal-mounted camera to take dense, sharp frames of the same scene. From the sharp frames, they reconstruct a 3D Gaussian representation (using 3D Gaussian Splatting) and then render a sharp image at the exact pose of each blurry frame.

The key innovation is the Blur-aware Pose Refinement (BPR) module, which refines the initial camera pose of the blurry image using appearance consistency and centroid alignment constraints. This ensures the rendered sharp counterpart aligns perfectly with the blurry image, avoiding misalignment artifacts common in real-world datasets. The team built a diverse dataset covering indoor and outdoor scenes with various motion types. Experiments show that a standard deblurring model (e.g., Restormer) trained on GS-RealBlur achieves superior generalization across multiple real-world benchmarks (e.g., RealBlur, GoPro), beating models trained on synthetic datasets by up to 0.5 dB PSNR and outperforming existing real-world datasets. The framework is flexible: the gimbal can be replaced with any device that captures sharp images, and the 3D reconstruction works even with sparse input. Code and dataset will be public, lowering the barrier for high-quality deblurring research.