Assessing the Role of Datasets in the Generalization of Motion Deblurring Methods to Real Images

TL;DR

This work addresses the generalization gap of motion deblurring methods to real images by diagnosing dataset-related limitations and proposing a scalable, segmentation-based non-uniform blur synthesis method (SBDD). By modeling per-object blur kernels, saturation, and realistic CRFs in the photon domain, the authors generate a virtually unlimited variety of training pairs and demonstrate improved cross-dataset generalization across GoPro, Köhler, RealBlur, and Lai datasets. Experiments show that non-uniform blur training enhances performance on dynamic scenes, while uniform blur training remains effective for smoothly varying blur; CRF awareness and saturation handling are critical for real-world restoration. The results indicate a data-centric path to boosting real-image deblurring performance across architectures, with code and dataset generation details made publicly available for broad adoption.

Abstract

Successfully training end-to-end deep networks for real motion deblurring requires datasets of sharp/blurred image pairs that are realistic and diverse enough to achieve generalization to real blurred images. Obtaining such datasets remains a challenging task. In this paper, we first review the limitations of existing deblurring benchmark datasets and analyze the underlying causes for deblurring networks' lack of generalization to blurry images in the wild. Based on this analysis, we propose an efficient procedural methodology to generate sharp/blurred image pairs based on a simple yet effective model. This allows for generating virtually unlimited diverse training pairs mimicking realistic blur properties. We demonstrate the effectiveness of the proposed dataset by training existing deblurring architectures on the simulated pairs and performing cross-dataset evaluation on three standard datasets of real blurred images. When training with the proposed method, we observed superior generalization performance for the ultimate task of deblurring real motion-blurred photos of dynamic scenes.

Figures (14)

• Figure 1: PSNR performance in the widely used GoPro dataset Nah_2017_CVPR (represented by circles diameters) is not indicative of performance in real blurred photos (Köhler and RealBlur datasets). When state-of-the art (MIMO-UNet and NAFNet) and classic (SRN) deblurring networks are trained using the GoPro dataset sharp images, but synthesizing the blurry images with the proposed procedure, the methods generalize better (filled circles to hollow circles transitions).
• Figure 2: State-of-the-art deblurring neural networks achieve spectacular restorations whithin the dataset they are trained on, but generalize poorly to real blurred images. We conjecture that this is due to a discrepancy between the training set underlying degradation model and that of the actual blurred photographs of dynamic scenes. We propose a new model-based methodology for generating training pairs, that improves model performance in the ultimate task of deblurring real blurred images (right column)
• Figure 3: A synthetic pattern with varying height steps was generated to show the effect of the Camera Response Function (CRF) on the blurry borders of an image. The CRF is modeled as $g({x})={x}^{1/\gamma}$. As the gamma value increases, the light steps look wider than the dark ones.
• Figure 4: Blurry/sharp pair crop sourced from the GoPro dataset for the linear and $\gamma$-corrected cases. The crops correspond to a saturated case. The blue arrows indicate the ghosting effect, while the green arrows show that the $\gamma$-corrected image has "whiter" transitions. The black arrow indicates the door opener borders present in the synthetic blurry images. However, these borders would not have been visible in a real blurry image since the surrounding pixels are saturated. Best viewed in electronic format.
• Figure 5: Blurry/sharp pair crop sourced from the RealBlur dataset It is observed that saturated regions in the blurry image are larger, with higher pixel values than the corresponding sharp image, contradicting the motion blur model. Best viewed in electronic format.