Multi-Scale Separable Network for Ultra-High-Definition Video Deblurring

Deng, Senyou; Ren, Wenqi; Yan, Yanyang; Wang, Tao; Song, Fenglong; Cao, Xiaochun

doi:10.1109/iccv48922.2021.01377

Cited by 21 publications

(16 citation statements)

References 42 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…HIDE [32] provides synthetic blurred images with humans and bounding box labels of humans. To expand deblurring into high-resolution images, 4KRD [12] is presented, which consists of synthetically blurred UHD video frames. All the datasets discussed above except for GoPro use frame interpolation before averaging sharp images to synthesize more realistic blur [12,26,32,33,48].…”

Section: Related Workmentioning

confidence: 99%

“…To expand deblurring into high-resolution images, 4KRD [12] is presented, which consists of synthetically blurred UHD video frames. All the datasets discussed above except for GoPro use frame interpolation before averaging sharp images to synthesize more realistic blur [12,26,32,33,48]. HFR-DVD [22] uses high-speed video frames captured at 1000 FPS to synthesize blur without frame interpolation.…”

Section: Related Workmentioning

confidence: 99%

“…In the denoising field, it has been proven important to model realistic noise for high-quality denoising of real-world images [1,4,7,15,36,44]. On the other hand, noise is ignored by the blur generation processes of the previous synthetic datasets [12,22,25,26,32,33,48], and its effect on deblurring has not been investigated.…”

Section: Discontinuous Blur Trajectoriesmentioning

confidence: 99%

“…Camera ISP ISPs perform various operations including white balancing, color correction, demosaicing, and nonlinear mapping using camera response functions, which affect the noise distribution and introduce distortions [4,6]. However, they are ignored by the previous synthetic datasets [12,22,25,26,32,33,48].…”

Section: Discontinuous Blur Trajectoriesmentioning

confidence: 99%

“…Training learning-based deblurring methods demands a significant amount of blurred and sharp image pairs. Since it is hard to obtain real-world blurred and sharp image pairs, a number of synthetically generated datasets have been proposed, whose blurred images are generated by blending sharp video frames captured by high-speed cameras [12,22,25,26,32,33,48]. Unfortunately, such synthetic images are not realistic enough, so deblurring methods trained on them do not generalize well to real blurred images [30].…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

Realistic Blur Synthesis for Learning Image Deblurring

Rim¹,

Kim²,

Kim³

et al. 2022

Preprint

View full text Add to dashboard Cite

Training learning-based deblurring methods demands a significant amount of blurred and sharp image pairs. Unfortunately, existing synthetic datasets are not realistic enough, and existing real-world blur datasets provide limited diversity of scenes and camera settings. As a result, deblurring models trained on them still suffer from the lack of generalization ability for handling real blurred images. In this paper, we analyze various factors that introduce differences between real and synthetic blurred images, and present a novel blur synthesis pipeline that can synthesize more realistic blur. We also present RSBlur, a novel dataset that contains real blurred images and the corresponding sequences of sharp images. The RSBlur dataset can be used for generating synthetic blurred images to enable detailed analysis on the differences between real and synthetic blur. With our blur synthesis pipeline and RSBlur dataset, we reveal the effects of different factors in the blur synthesis. We also show that our synthesis method can improve the deblurring performance on real blurred images.

show abstract