Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Deblurring in the Wild: A Real-World Image Deblurring Dataset from Smartphone High-Speed Videos

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper introduces SloMoDeblur, a dataset of more than 42,000 blur-sharp image pairs built by averaging 30 consecutive frames from 240 fps smartphone videos, with the middle frame as the sharp reference.

desk verdict A large, genuinely useful smartphone-domain deblurring benchmark built with the GoPro frame-averaging recipe; the 'real-world' label is doing too much work, but the artifact is worth serious peer review. read the letter →

arxiv 2506.19445 v4 pith:6CS47XMH submitted 2025-06-24 cs.CV cs.AIcs.LG

Syed Mumtahin Mahmud , Mahdi Mohd Hossain Noki , Prothito Shovon Majumder , Abdul Mohaimen Al Radi , Sudipto Das Sukanto , Afia Lubaina , Md. Mosaddek Khan This is my paper
classification cs.CVcs.AIcs.LG
keywords imagedeblurringmotionblurdatasetsmartphonephotographyslow-motionvideoframeaveragingsingle-imagebenchmarkhigh-frame-rate
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that averaging 30 consecutive frames from a 240 fps smartphone video simulates realistic motion blur, while the window's middle frame serves as the sharp ground truth. It introduces SloMoDeblur, a dataset of more than 42,000 aligned blur-sharp pairs across 843 scenes, roughly ten times larger than GoPro and more than eleven times larger than RealBlur by pair count. The authors benchmark seven state-of-the-art deblurring models and find that none reaches the baseline PSNR of 32.38 dB set by comparing the blurry input directly with the sharp ground truth. The point of the work is to give the field a larger, smartphone-specific benchmark that better reflects the blur ordinary users experience. If the dataset holds up, it changes what counts as a hard deblurring test set and gives training a bigger, more diverse pool.

What carries the argument

The blur formation model $B' = \frac{1}{N}\sum_{i=0}^{N-1} F_i$ over 30 frames at 240 fps, giving an effective $1/8$-second exposure, with the center frame $F_{\mathrm{GT}} = F_{\lfloor N/2 \rfloor}$ as ground truth. This discrete approximation of the continuous integral $B = \frac{1}{T}\int_0^T F(t)\,dt$ carries the whole dataset construction, generating both the blurred image and the sharp reference from the same sensor at the same moment and thereby sidestepping the geometric alignment problems of dual-camera capture.

What would settle it

Shoot the same scene twice from a fixed position, once as a genuine one-eighth-second still exposure on the iPhone and once as the 30-frame average of a 240 fps video; if the averaged video frame differs strongly from the still exposure in edges, highlights, or overall pixel values, the blur-formation model does not reproduce physical long-exposure blur. Independently, align the center frame of each window against a tripod still of the same scene; visible residual motion blur in the center frame would mean the ground truth is not sharp.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that high-frame-rate smartphone video can be turned, by a simple temporal mean, into a large corpus of physically motivated blur-sharp pairs without specialized dual-camera hardware. The central finding is that current state-of-the-art deblurring models, when evaluated without fine-tuning, all score below the blurry-input baseline in PSNR on the test split, with SSIM improvements only marginal. That is, the dataset is measurably harder than existing benchmarks for the models tested, and no model in the evaluation set improves enough on the blurry input to justify the restoration effort in pixel-wise terms.

Load-bearing premise

The dataset's validity rests on the assumption that averaging 30 consecutive 240-frames-per-second video frames is physically equivalent to a real one-eighth-second exposure, and that the middle frame is a sharp, correctly aligned ground truth, which requires the phone's video signal to be linear in radiance and the motion in each window to be locally uniform.

Editorial extensions

If this is right

  • Training on this dataset should reduce the gap between benchmark performance and real smartphone photo deblurring, since the blur is generated from a consumer phone's actual video pipeline.
  • The 42,045 pairs allow modern data-hungry architectures such as transformers and state-space models to train from scratch rather than rely on synthetic kernels or small real datasets.
  • Because all models were tested without fine-tuning, the dataset provides a clean cross-dataset generalization benchmark; future work can report fine-tuned performance against these numbers.
  • The single-device design yields a stable, reproducible benchmark whose domain coverage is deliberately narrow, making extensions to low-light and multiple devices the next natural step.
  • The failure of state-of-the-art models to beat the blurry-input PSNR baseline on this test set suggests that the dataset captures harder, more realistic degradation than existing benchmarks.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural test would be to compare the averaged 30-frame blur with a true photograph taken at a $1/8$-second exposure on a tripod; where they disagree, the dataset inherits the video pipeline's tone-mapping and compression rather than true optical blur.
  • The same acquisition recipe could be extended to other smartphone models or paired with gyroscope data, turning the dataset into a motion-aware benchmark for physics-informed deblurring models.
  • If residual motion blur remains in the center frame, part of the measured PSNR gap could be due to imperfect ground truth rather than model weakness, which a perceptual study or deblurring of the ground-truth frames themselves could disentangle.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper introduces SloMoDeblur, a deblurring dataset of 42,045 blur-sharp image pairs generated from iPhone 15 Pro 240 fps videos by averaging 30 consecutive frames to synthesize blur and using the middle frame as the sharp ground truth (Sec. 3, Eqs. 1-3). The authors benchmark several recent deblurring models on a 10% test split and report PSNR/SSIM around or below the blurry-vs-GT baseline, interpreting this as evidence of dataset difficulty. The manuscript claims the dataset is the largest of its kind and more diverse than GoPro/RealBlur, and argues that smartphone-domain blur better represents consumer photography. The central physical assumption — that temporal averaging of video frames equals a real long-exposure blur — is asserted rather than validated.

Significance. If validated, the dataset would be a valuable resource: its scale (42,045 pairs) and smartphone domain are genuinely useful for training and benchmarking, and the release of generation scripts and use of a reproducible pipeline are strengths. However, the contribution's load-bearing claim is 'real-world' fidelity, and the current evidence for physical correctness is missing. The paper itself labels the data 'Synth. (phone)' (Table 2), and Section 3.4's linear-RGB assertion is unsubstantiated. The benchmark results are also presented without error bars and with only pretrained checkpoints, limiting the strength of the difficulty claim. These issues are fixable with additional validation experiments and reporting.

major comments (5)
  1. [Sec. 3.4 and Eq. (2)] The dataset's realism rests on the claim that averaging 30 video frames yields the same image as a true 1/8 s exposure, which requires the video frames to be linear in scene radiance. The paper asserts that 'we extract linear RGB data directly from the iPhone's video stream' but provides no CRF calibration, no comparison with linear raw frames, and no validation against a real long exposure. This is load-bearing: if the frames are tone-mapped or compressed, the temporal average is not a physical motion blur and the ground-truth center frame is not a clean latent image. Please add a radiometric calibration or a direct validation experiment, or scale back the 'real-world' claim to 'smartphone-domain synthetic blur.'
  2. [Sec. 3.2, Eq. (3)] The choice of the center frame as sharp ground truth assumes that motion within the 30-frame window is negligible for that frame, but no evidence of sharpness is provided. For fast-moving objects or camera shake, the center frame may itself contain motion blur, making the 'sharp' reference blurry. Please quantify residual blur (e.g., edge width statistics, comparison against a shorter-exposure reference, or per-pixel optical-flow magnitude) and report how many pairs have potentially non-negligible center-frame blur.
  3. [Abstract and Sec. 3.4 vs. Table 2] The paper interchangeably calls the dataset 'real-world' (abstract, conclusion) and 'synthetic' ('Synth. (phone)' in Table 2; 'synthetically generated via frame averaging' in Section 4.1). This is not a terminological nit: the central contribution is framed as a real-world benchmark, and the reader cannot tell which claim is being made. Please adopt consistent terminology and define precisely what 'real-world' means (real scenes and motions vs. physically captured long-exposure blur).
  4. [Sec. 3.4 vs. Sec. 3] The paper claims the dataset includes 'linear, rotational, and occlusive dynamics, as well as non-uniform motion blur,' yet Section 3 states that the synthesis 'assumes approximately uniform motion within the averaging window and does not fully reproduce more complex patterns of blur arising from non-linear motion trajectories or occlusions.' These statements are in direct tension. If the dataset contains such complex blur types, please demonstrate this with quantitative motion analysis; otherwise remove the claim, since Eq. (2) only models uniform-motion blur.
  5. [Sec. 4.1, Table 1] The benchmark results have no error bars or statistical significance tests, and the models are evaluated with publicly available pretrained weights (acknowledged in Sec. 5.1) rather than trained on the proposed train split. The headline finding that models 'perform below the PSNR baseline' may reflect domain shift from the pretraining data rather than intrinsic dataset difficulty. Please report variance over the test split, and preferably include a model trained on the dataset's training split, to separate domain shift from benchmark difficulty.
minor comments (6)
  1. [Abstract] The sentence 'We release the dataset and generation scripts athuggingface.' is a formatting fragment; please include the full URL or Hugging Face repository identifier.
  2. [Abstract and Sec. 1] The claim of '8 times the amount of different scenes' is ambiguous: 843/110 ≈ 7.7x only versus RealBlur, while versus GoPro it is about 2.5x; please state the comparison base explicitly.
  3. [Algorithm 1] The indentation of lines 10-13 is ambiguous; as printed, 'blurred' and the save operations appear inside the for loop, and the ground-truth frame is saved inside the if block rather than after the accumulation loop. Please clarify the control flow.
  4. [Sec. 2.1] The statement that state-of-the-art deblurring methods achieve PSNR scores above 34 dB and SSIM close to 0.97 is uncited; please add references for these numbers.
  5. [Sec. 3.4] The phrase 'linear RGB data directly from the iPhone's video stream' is inconsistent with the same section's claim that the dataset 'inherently incorporates ... tone-mapping behavior and in-camera processing'; please clarify which processing steps are applied.
  6. [Table 2] The row for Lai et al. reports 100 blurry images but the dataset has no sharp ground truth; the table's 'Type' column should note 'unpaired,' as the text does, to avoid implying it contains blur-sharp pairs.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the blur and sharp pairs are generated by explicit construction (Eqs. 2–3), benchmark results are measured rather than fitted, and no load-bearing self-citation appears.

full rationale

The paper's core artifact is a dataset, not a derived physical law. The blurry image is defined in Eq. (2) as the mean of 30 video frames and the ground truth is defined in Eq. (3) as the central frame; these are explicit data-generation operations, and the paper does not claim to predict one from the other. The benchmark experiments measure PSNR/SSIM of publicly available pretrained models on a fixed 10% test split, with no fitted parameters or model-specific tuning aimed at forcing the reported degradation; the 'Blurry vs. GT' baseline is a direct comparison of two constructed images, not a circular prediction. The paper discloses that the dataset is labeled 'Synth. (phone)' in Table 2, honestly acknowledging that the blur is synthesized by frame averaging rather than captured as a true long exposure. The main validity concern, namely whether iPhone 240 fps frames are radiometrically linear so that averaging them equals physical motion blur, is an empirical calibration or validation gap; it is not a circularity because the paper never uses Eq. (2) to establish that linearity, nor does it fit any parameter to make the benchmark appear more challenging. There are also no load-bearing self-citations: all cited datasets and methods are external prior work, and the GoPro-style averaging methodology is attributed to earlier public datasets. Accordingly, the derivation chain is self-contained and no circular step can be exhibited.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The dataset generation relies on one hand-chosen averaging window and two domain assumptions: video frames average like linear radiance, and motion within the window is uniform enough that the center frame is sharp. No new physical entities are introduced.

free parameters (1)
  • Averaging window size N = 30 frames (1/8 second)
    Chosen by hand as a trade-off between motion realism and photometric integrity. Section 3.1 states 'We fixed N=30', and Section 3.4 explains it was selected empirically rather than derived from data.
assumptions (2)
  • domain assumption Frames within a 30-frame window are temporally aligned and motion is approximately uniform, so the center frame approximates a sharp latent image.
    Section 3.2: 'We assume the objects are in uniform motion with the very small temporal window, and with 240 fps, the center frame approximates a latent sharp image with negligible distortion.' This underpins the validity of both ground truth and blur synthesis.
  • domain assumption The smartphone video pipeline outputs approximately linear RGB radiance values, so averaging pixel values mimics physical long-exposure integration.
    Section 3.4 states 'we extract linear RGB data directly from the iPhone’s video stream' with no characterization of tone mapping or compression. Equation (2) averages pixel values directly, which is only physically correct under linear response.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deblurring in the Wild: A Real-World Image Deblurring Dataset from Smartphone High-Speed Videos." pith.science (2026). https://pith.science/paper/6CS47XMH

@misc{pith2026250619445,
  author       = {Pith},
  title        = {Pith review of: Deblurring in the Wild: A Real-World Image Deblurring Dataset from Smartphone High-Speed Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6CS47XMH}},
  note         = {Machine review of arXiv:2506.19445}
}
read the original abstract

We introduce the largest real-world image deblurring dataset constructed from smartphone slow-motion videos. Using 240 frames captured over one second, we simulate realistic long-exposure blur by averaging frames to produce blurry images, while using the temporally centered frame as the sharp reference. Our dataset contains over 42,000 high-resolution blur-sharp image pairs, making it approximately 10 times larger than widely used datasets, with 8 times the amount of different scenes, including indoor and outdoor environments, with varying object and camera motions. We benchmark multiple state-of-the-art (SOTA) deblurring models on our dataset and observe significant performance degradation, highlighting the complexity and diversity of our benchmark. Our dataset serves as a challenging new benchmark to facilitate robust and generalizable deblurring models.

Figures

Figures reproduced from arXiv: 2506.19445 by the authors.

Figure 1
Figure 1. Comparison of SloMoDeblur dataset with other popular datasets. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. (a) Extracting pixel data from an image frame (b) Layering temporally close frames to produce a synthetic blur [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Paired images with sharp and blurred versions. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. RealVDeblur: One-Step Diffusion for Generalizable Real-World Video Deblurring

    cs.CV 2026-07 conditional novelty 6.0 of 10

    RealVDeblur trains a one-step video-diffusion deblurrer on a large 3DGS-based synthetic dataset and stabilizes long-video inference with a temporal window mask, improving perceptual quality on real-world benchmarks.

Reference graph

Works this paper leans on

33 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    A neural approach to blind motion deblur- ring

    Ayan Chakrabarti. A neural approach to blind motion deblur- ring. 2016. 1

    work page 2016

  2. [2]

    Simple baselines for image restoration.arXiv preprint arXiv:2204.04676, 2022

    Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration.arXiv preprint arXiv:2204.04676, 2022. 6

    arXiv 2022

  3. [3]

    Fast motion deblurring

    Sunghyun Cho and Seungyong Lee. Fast motion deblurring. ACM Trans. Graph., 28, 2009. 1, 2

    work page 2009

  4. [4]

    Rethinking coarse-to-fine approach in single image deblurring

    Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 4621–4630,

    work page

  5. [5]

    Image restoration via frequency selection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023

    Yuning Cui, Wenqi Ren, Xiaochun Cao, and Alois Knoll. Image restoration via frequency selection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2

    work page 2023

  6. [6]

    Roweis, and William T

    Rob Fergus, Barun Singh, Aaron Hertzmann, Sam T. Roweis, and William T. Freeman. Removing camera shake from a single photograph.ACM Trans. Graph., 25(3):787–794, 2006. 1, 2

    work page 2006

  7. [7]

    Zitnick, Michael Cohen, and Brian Curless

    Ankit Gupta, Neel Joshi, C. Zitnick, Michael Cohen, and Brian Curless. Single image deblurring using motion density functions. pages 171–184, 2010

    work page 2010

  8. [8]

    Space-variant single-image blind deconvolution for removing camera shake

    Stefan Harmeling, Hirsch Michael, and Bernhard Schölkopf. Space-variant single-image blind deconvolution for removing camera shake. InAdvances in Neural Information Processing Systems. Curran Associates, Inc., 2010

    work page 2010

Show all 33 references
  1. [9]

    Schuler, Stefan Harmeling, and Bernhard Schölkopf

    Michael Hirsch, Christian J. Schuler, Stefan Harmeling, and Bernhard Schölkopf. Fast removal of non-uniform camera shake. In2011 International Conference on Computer Vision, pages 463–470, 2011. 1, 2

    2011

  2. [10]

    Mohler, Bernhard Scholkopf, and Stefan Harmeling

    Rolf Köhler, Michael Hirsch, Betty J. Mohler, Bernhard Scholkopf, and Stefan Harmeling. Recording and playback of camera shake: Benchmarking blind deconvolution with a real-world database. InEuropean Conference on Computer Vision, 2012. 2, 6, 7

    2012

  3. [11]

    Efficient frequency domain-based trans- formers for high-quality image deblurring

    Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based trans- formers for high-quality image deblurring. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5886–5895, 2023. 2, 6

    2023

  4. [12]

    Efficient visual state space model for image deblur- ring, 2024

    Lingshun Kong, Jiangxin Dong, Ming-Hsuan Yang, and Jin- shan Pan. Efficient visual state space model for image deblur- ring, 2024. 2

    2024

  5. [13]

    A Comparative Study for Single Image Blind Deblurring

    Wei-Sheng Lai, Jia-Bin Huang, Zhe Hu, Narendra Ahuja, and Ming-Hsuan Yang. A Comparative Study for Single Image Blind Deblurring . In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1701–1709, Los Alamitos, CA, USA, 2016. IEEE Computer Society. 2, 7

    2016

  6. [14]

    Free- man

    Anat Levin, Yair Weiss, Fredo Durand, and William T. Free- man. Understanding and evaluating blind deconvolution algo- rithms. In2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 1964–1971, 2009. 2, 6, 7 8

    1964

  7. [15]

    Intriguing findings of frequency selection for image deblurring.Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):1905–1913, 2023

    Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring.Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):1905–1913, 2023. 2

    1905

  8. [16]

    Adarevd: Adaptive patch exiting reversible decoder pushes the limit of image deblurring

    Xintian Mao, Qingli Li, and Yan Wang. Adarevd: Adaptive patch exiting reversible decoder pushes the limit of image deblurring. pages 25681–25690, 2024. 2, 6

    2024

  9. [17]

    Loformer: Local frequency transformer for image deblurring, 2024

    Xintian Mao, Jiansheng Wang, Xingran Xie, Qingli Li, and Yan Wang. Loformer: Local frequency transformer for image deblurring, 2024. 2

    2024

  10. [18]

    Deep multi-scale convolutional neural network for dynamic scene deblurring

    Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. InCVPR, 2017. 2, 7

    2017

  11. [19]

    Deblurring images via dark channel prior.IEEE Trans

    Jinshan Pan, Deqing Sun, Hanspeter Pfister, and Ming-Hsuan Yang. Deblurring images via dark channel prior.IEEE Trans. Pattern Anal. Mach. Intell., 40(10):2315–2328, 2018. 1, 2

    2018

  12. [20]

    Real-world blur dataset for learning and benchmarking de- blurring algorithms

    Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking de- blurring algorithms. InComputer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXV, page 184–201, Berlin, Heidelberg,

    2020

  13. [21]

    Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf

    Christian J. Schuler, Michael Hirsch, Stefan Harmeling, and Bernhard Schölkopf. Learning to deblur.IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(7):1439– 1451, 2016. 1, 2

    2016

  14. [22]

    High-quality motion deblurring from a single image.ACM Trans

    Qi Shan, Jiaya Jia, and Aseem Agarwala. High-quality motion deblurring from a single image.ACM Trans. Graph., 27(3): 1–10, 2008. 1, 2

    2008

  15. [23]

    Human-aware motion deblur- ring

    Ziyi Shen, Wenguan Wang, Jianbing Shen, Haibin Ling, Tingfa Xu, and Ling Shao. Human-aware motion deblur- ring. InIEEE International Conference on Computer Vision,

  16. [24]

    Learn- ing a convolutional neural network for non-uniform motion blur removal

    Jian Sun, Wenfei Cao, Zongben Xu, and Jean Ponce. Learn- ing a convolutional neural network for non-uniform motion blur removal . In2015 IEEE Conference on Computer Vi- sion and Pattern Recognition (CVPR), pages 769–777, Los Alamitos, CA, USA, 2015. IEEE Computer Society. 1

    2015

  17. [25]

    Edge- based blur kernel estimation using patch priors

    Libin Sun, Sunghyun Cho, Jue Wang, and James Hays. Edge- based blur kernel estimation using patch priors. InIEEE Inter- national Conference on Computational Photography (ICCP), pages 1–8, 2013. 2, 7

    2013

  18. [26]

    Oliver Whyte, Josef Sivic, Andrew Zisserman, and J. Ponce. Non-uniform deblurring for shaken images. pages 491–498,

  19. [27]

    Two-phase kernel estimation for ro- bust motion deblurring

    Li Xu and Jiaya Jia. Two-phase kernel estimation for ro- bust motion deblurring. InComputer Vision – ECCV 2010, pages 157–170, Berlin, Heidelberg, 2010. Springer Berlin Heidelberg

    2010

  20. [28]

    Unnatural l0 sparse representation for natural image deblurring

    Li Xu, Shicheng Zheng, and Jiaya Jia. Unnatural l0 sparse representation for natural image deblurring. In2013 IEEE Conference on Computer Vision and Pattern Recognition, pages 1107–1114, 2013. 1, 2

    2013

  21. [29]

    Li Xu, Jimmy S. J. Ren, Ce Liu, and Jiaya Jia. Deep con- volutional neural network for image deconvolution. InPro- ceedings of the 28th International Conference on Neural In- formation Processing Systems - Volume 1, page 1790–1798, Cambridge, MA, USA, 2014. MIT Press. 1

    2014

  22. [30]

    Multi-Stage Progressive Image Restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-Stage Progressive Image Restoration . In 2021 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 14816–14826, Los Alamitos, CA, USA...

    2021

  23. [31]

    Restormer: Efficient transformer for high-resolution image restoration

    Syed Waqas Zamir, Aditya Arora, Salman Khan, Mu- nawar Hayat, Fahad Shahbaz Khan, and Ming–Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5718–5729,

  24. [32]

    Deep stacked hierarchical multi-patch network for image deblurring

    Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for image deblurring. InThe IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2, 6 9

    2019

  25. [2020]

    Springer-Verlag. 2, 7

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.