Pith. sign in

REVIEW 4 major objections 5 minor 43 references

PMR: Physical Model-Driven Multi-Stage Restoration of Turbulent Dynamic Videos

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A three-stage physical-model pipeline yields sharper restoration of turbulent dynamic videos, outperforming prior methods on high-dynamic scenes.

desk verdict Solid three-stage turbulence-restoration engineering, but the main quantitative claim rests on an unstated data split and an under-specified metric; worth refereeing with a demand for the missing split and code. read the letter →

arxiv 2508.00406 v1 pith:VQNEJTJE submitted 2025-08-01 cs.CV

classification cs.CV
keywords atmosphericturbulencevideorestorationmulti-stagenetworkdynamicefficiencyindexopticalflowdeblurringgeometricdistortiondeeplearning
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 sets out to establish that atmospheric-turbulence degradation of dynamic videos is best undone in three ordered steps—de-tilting, motion-segmentation enhancement, and de-blurring—mirroring a physical model in which turbulent tilt, object motion, and higher-order blur corrupt a clean frame in sequence. To make this work where prior methods fail, the authors introduce the Dynamic Efficiency Index (DEI), a scalar that combines turbulence strength, optical-flow magnitude, and the fraction of dynamic pixels, and use it to select a high-dynamic-intensity training subset. They report that the best variant, PMR_T, reaches 28.223 dB PSNR on the High DEI synthetic benchmark versus 27.853 dB for the previous best method, and that the three variants yield visually sharper edges and fewer motion-trailing artifacts on real-world 2K turbulent clips. The practical stake is that long-range imaging under strong turbulence—surveillance, remote sensing, outdoor observation—could get clearer, stabler video without paying for heavy transformer models.

What carries the argument

The load-bearing identity is the physical decomposition $I = T(M(B(J)))$ (Eq. 1), which dictates the stage order: de-tilting, motion-segmentation enhancement, de-blurring. The de-tilting network DET averages multi-scale tilt fields, exploiting the zero-mean Gaussian property of tilt to get a stable correction; MSE_OF uses RAFT optical flow with a $1/(1+C_n^2)$ attenuation plus Gaussian background weighting; DEB is a lightweight encoder-decoder whose STC_transformer blocks model spatio-temporal channel attention mostly at low resolution. The Dynamic Efficiency Index (DEI, Eq. 6) is the selection mechanism: it combines turbulence intensity $C_n^2$, binarized optical-flow magnitude, and the dynamic-region proportion through a hand-set piecewise coefficient $C$ and an undefined constant $\gamma$, and its threshold DEI=100 carves out the High DEI training subset.

What would settle it

Train the same PMR pipeline on a random subset of 1,568 synthetic videos matched in size to the High DEI subset and compare on the High DEI test set; if the random subset matches or beats DEI-selected training in PSNR, the index's contribution to the result is selection bias rather than dynamic-intensity measurement. Also, re-computing DEI with different values of $\gamma$ or $C$ and checking whether the performance gain disappears would test the index's robustness.

Watch

Extended reading notes

Core claim

The central claim is that a turbulent dynamic frame can be modeled as $I = T(M(B(J)))$: tilt distortion $T$, object-motion artifacts $M$, and blur $B$ applied in that order to a clean frame $J$, and that stripping them in reverse order stage by stage recovers the scene better than any single-stage or collapsed end-to-end approach. The paper supports this with a three-stage pipeline—DET for geometric stabilization via multi-scale tilt fields, MSE_OF for optical-flow-based foreground/background separation with turbulence-adaptive Gaussian smoothing, and DEB for residual deblurring with a hybrid CNN plus spatio-temporal channel attention. To feed this pipeline the hardest cases, the Dynamic Efficiency Index (DEI) is defined from $C_n^2$, adjusted optical flow, and dynamic-region proportion; videos with DEI ≥ 100 form the High DEI training set. On that benchmark the best variant PMR_T scores 28.223 dB PSNR and 0.845 SSIM, ahead of the 27.853 dB of TMT, and all PMR variants show reduced motion trailing and sharper edges on real 2K clips.

Load-bearing premise

The Dynamic Efficiency Index (DEI) truly measures dynamic intensity, and the chosen threshold DEI=100 selects a training subset that actually improves high-dynamic restoration; the paper does not validate DEI against any external ground-truth dynamic-intensity measure.

Editorial extensions

If this is right

  • Training on a DEI-selected high-dynamic subset improves restoration for all compared methods, not only PMR; if reproduced, this selection protocol could become standard practice in turbulence-restoration benchmarks.
  • The validated three-stage order (tilt → motion separation → blur removal) implies that future methods should not collapse these distortions into one monolithic end-to-end pass.
  • The lightweight DET and DEB modules (28.81M parameters) make 2K turbulence restoration feasible at about half the inference time of TMT, a step toward practical deployment.
  • Swapping the DEB backbone for stronger transformers (PMR_T, PMR_R) preserves the per-stage gains, showing the framework is backbone-agnostic and benefits from more powerful final-stage models.

Reading between the lines

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

  • A testable extension is to apply the same DEI-style dynamic-intensity data selection to other video restoration tasks (rain, snow, haze) to see whether the selection effect generalizes beyond atmospheric turbulence.
  • The DEI threshold and its internal constants ($\gamma$, piecewise $C$) were hand-chosen after experiments; an automated calibration that optimizes downstream restoration PSNR could either strengthen or weaken the index's role.
  • The de-tilting stage relies on averaging tilt fields over time under a zero-mean Gaussian assumption; if real turbulence exhibits non-stationary bursts, the same averaging could smear fast-moving objects, an effect worth testing on annotated real-motion clips.
  • The real-world generalization claim rests on visual comparisons; a perceptual study or a no-reference metric on the real 2K clips would quantify the reported sharpness advantage.
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

4 major / 5 minor

Summary. The paper proposes PMR, a three-stage pipeline for restoring videos degraded by atmospheric turbulence and dynamics: de-tilting for geometric stabilization, motion-segmentation enhancement for dynamic regions, and de-blurring for local blur. It also introduces a Dynamic Efficiency Index (DEI) to quantify the dynamic intensity of turbulent videos and to construct a high-dynamic training subset. The authors report that PMR_T achieves the best PSNR (28.223 dB) on the High DEI test set and show qualitative improvements on real 2K turbulent clips. The paper claims strong generalization, especially in high-turbulence, complex-motion scenarios, and promises to release code and datasets.

Significance. If the claims hold, the paper would make a useful contribution to turbulence-robust video restoration: it offers a physically motivated decomposition of tilt, motion, and blur; introduces a lightweight multi-stage architecture; and proposes a metric (DEI) for selecting dynamic training data. The quantitative gains on the high-dynamic synthetic set over methods such as TMT are modest but consistent across PSNR, SSIM, and LPIPS, and the qualitative results on real-world 2K videos are compelling. The paper also provides ablation studies on its lightweight modules and training strategy, and promises open code and data, which are positive features. However, the centrality of DEI to both dataset construction and evaluation, together with unspecified dataset partitioning and undefined constants, creates a correctness risk that must be resolved before the main claims can be accepted.

major comments (4)
  1. [Sections 4.1 and 4.2] The paper does not specify whether the 713 High DEI test videos are disjoint from the 1,568 videos used for DEI-selected feature-enhanced training. Section 4.1 states that 1,568 of 3,500 synthetic videos were selected for training using a DEI=100 threshold, and Section 4.2 reports evaluation on 713 High DEI videos. The numbers 1,568 + 1,184 + 713 = 3,465 leave 35 videos unaccounted for, and no train/test split is described. If any test videos were used for training, the headline result of 28.223 dB in Table 1 is not an independent measure of generalization. The authors must release exact video indices, the partition rule, and confirm disjointness.
  2. [Section 3.1, Eq. (4)] Equation (4) is mathematically inconsistent as written. DyOF_i is defined as Max(davg(Norm(OF_i), 0.5)), which appears to operate on normalized optical flow, while the denominator in DPR is the sum of raw optical flow OF_j. Mixing normalized and unnormalized quantities makes DPR dependent on an arbitrary scaling and not reproducible. Please define the normalization, the binarization threshold, and the exact operations in Eq. (4) so that DPR is a well-defined, scale-invariant quantity, or explain why the denominator is intentionally unnormalized.
  3. [Section 3.1, Eq. (6)] The DEI definition depends on an undefined constant gamma and a hand-picked piecewise coefficient C from Eq. (5). The paper does not specify the value of gamma, the source or justification for the thresholds in Eq. (5), or how these choices affect the claimed threshold of DEI=100. Since DEI is used both to select the training subset and to define the High DEI evaluation set, these unspecified parameters make the central evaluation protocol non-reproducible. The authors should provide the exact formula with tuned constants and, ideally, a sensitivity analysis over gamma and the C thresholds.
  4. [Section 4.1 and Table 1] It is unclear whether competing methods (TMT, TurbSR, Deturb, VRT, etc.) were trained on the same high-DEI subset as PMR. Section 4.1 describes the DEI-selected training data for PMR and says that other models use the same optimizer and scheduler, but it does not state whether those models also received the 1,568-video high-DEI training set. If competitors were trained only on the original Normal dataset while PMR was trained on a high-dynamic subset, Table 1 may reflect a training-data advantage rather than architectural superiority. Please specify the exact training data for every evaluated method.
minor comments (5)
  1. [Section 3.2.1] In the text after Figure 3, 'a more table tilt field' appears to be a typo for 'stable tilt field'.
  2. [Section 3.2.1, Eq. (7)] The function 'Harr' presumably denotes the Haar wavelet; please correct the spelling and cite the wavelet used.
  3. [References] References [29] and [30] are duplicate entries for the same RAFT paper; one should be removed or the citations disambiguated.
  4. [Section 4.2, Table 3 discussion] The paragraph after Table 3 contains an incomplete sentence: 'Single-stage methods such as VRT, although delivering excellent restoration performance, suffer from extremely high inference time and computational cost, size, making it less suitable for real-time deployment.' The phrase appears truncated and should be rewritten.
  5. [Section 3.2.2, Eq. (8)] The definition of OFD is difficult to parse: the notation with nested max/mean and absolute values should be expanded into a clearer mathematical expression or a verbal explanation of the intended 'uniformity' measure.

Circularity Check

0 steps flagged · score 2.0 of 10

No demonstrated circularity; DEI train/test protocol is under-specified but not a proven circular step.

full rationale

The paper's central derivation chain is the physical model Eq. (1), which decomposes a turbulent dynamic frame into tilt (T), motion (M), and blur (B). The PMR framework inverts this decomposition with three stages (de-tilting, motion-segmentation enhancement, de-blurring), each targeting one component. This is a modeling choice guided by physics, not a circular step: the architecture mirrors the decomposition by design, and the components are defined independently of the network. The DEI metric (Eqs. 2-6) is an empirical index used to select a feature-enhanced training subset (Section 4.1: 'we selected 1,568 videos as the high subset DEI for feature-enhanced training, using DEI = 100 as the threshold after multiple experiments and visual analysis') and to name a test set (Section 4.2: 'the High DEI dataset includes 713 videos'). This creates a potential self-referential evaluation protocol: the same index defines both train and test distributions, and the paper does not explicitly state that the 713 test videos are disjoint from the 1,568 training videos. The arithmetic (1,568 + 1,184 + 713 = 3,465 out of 3,500) leaves 35 videos unaccounted for, which suggests disjointness but does not prove it. If overlap exists, the reported High-DEI PSNR would be training-set leakage; however, the paper does not claim overlap, and the evaluation also includes the external Normal dataset and real-world 2K clips, which provide independent evidence. The undefined constant gamma in Eq. (6) and the hand-picked coefficient C in Eq. (5) make DEI non-reproducible, but these are correctness/reproducibility issues, not circular reductions. The paper cites prior work by the same authors (e.g., refs [15], [21], [24]) as baselines or background, but these citations are not load-bearing for the central claim; the SOTA comparison is against external methods. Overall, the derivation does not reduce to its inputs by construction, and no specific circular step can be exhibited from the paper's own equations.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central claim rests on several hand-set constants: the DEI threshold of 100 (Section 4.1), the piecewise C coefficients in Eq. (5), the optical-flow binarization threshold 0.5 in Eq. (4), and the undefined turbulence constant gamma in Eq. (6). The physical decomposition T(M(B(J))) in Eq. (1) is asserted, not derived, and the reliability of RAFT optical flow under turbulence is a domain assumption the paper itself flags as problematic. These are not fatal, but they mean the reported performance gains depend on choices that are only partially justified.

free parameters (4)
  • DEI threshold = 100
    Used to select the High DEI training subset from the 3,500 synthetic videos; chosen 'after multiple experiments and visual analysis' (Section 4.1).
  • DPR piecewise thresholds and coefficients (C) = thresholds 0.05, 0.5, 0.7; coefficients 2, 1, 0.5, 0.1
    Eq. (5) defines C with hand-selected breakpoints; no justification beyond 'experimental results indicate' (Section 3.1).
  • Optical-flow binarization threshold = 0.5
    Eq. (4) uses d_avg from threshold 0.5 to define dynamic pixels; no derivation.
  • Turbulence constant gamma = not defined
    Eq. (6) divides by gamma without defining it; the value is needed to compute DEI.
assumptions (4)
  • standard math Tilt displacement follows a zero-mean Gaussian process
    Cited from Fried [4] and prior turbulence models; standard in the field.
  • ad hoc to paper Turbulence distortion can be decomposed as T(M(B(J))) in that order
    Eq. (1) states the decomposition without derivation; the multi-stage architecture directly mirrors it.
  • domain assumption Optical flow from RAFT is reliable for motion segmentation after de-tilting
    The MSE_OF stage relies on RAFT flow; the paper itself notes turbulence violates brightness constancy, and only applies a scaling factor 1/(1+C_n^2).
  • domain assumption The synthetic dataset from [39] is representative of real turbulent dynamic scenes
    Training and most evaluation use synthetic data; real-world test is only qualitative.
invented entities (1)
  • Dynamic Efficiency Index (DEI)
    purpose: Quantify dynamic intensity of turbulent videos and curate a high-dynamic training subset
    No external validation of DEI against ground-truth dynamic intensity; it is defined using hand-picked thresholds and an undefined gamma.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PMR: Physical Model-Driven Multi-Stage Restoration of Turbulent Dynamic Videos." pith.science (2026). https://pith.science/paper/VQNEJTJE

@misc{pith2026250800406,
  author       = {Pith},
  title        = {Pith review of: PMR: Physical Model-Driven Multi-Stage Restoration of Turbulent Dynamic Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VQNEJTJE}},
  note         = {Machine review of arXiv:2508.00406}
}
abstract

Geometric distortions and blurring caused by atmospheric turbulence degrade the quality of long-range dynamic scene videos. Existing methods struggle with restoring edge details and eliminating mixed distortions, especially under conditions of strong turbulence and complex dynamics. To address these challenges, we introduce a Dynamic Efficiency Index ($DEI$), which combines turbulence intensity, optical flow, and proportions of dynamic regions to accurately quantify video dynamic intensity under varying turbulence conditions and provide a high-dynamic turbulence training dataset. Additionally, we propose a Physical Model-Driven Multi-Stage Video Restoration ($PMR$) framework that consists of three stages: \textbf{de-tilting} for geometric stabilization, \textbf{motion segmentation enhancement} for dynamic region refinement, and \textbf{de-blurring} for quality restoration. $PMR$ employs lightweight backbones and stage-wise joint training to ensure both efficiency and high restoration quality. Experimental results demonstrate that the proposed method effectively suppresses motion trailing artifacts, restores edge details and exhibits strong generalization capability, especially in real-world scenarios characterized by high-turbulence and complex dynamics. We will make the code and datasets openly available.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 27 canonical work pages

  1. [1]

    Anantrasirichai

    N. Anantrasirichai. Atmospheric turbulence removal with complex- valued convolutional neural network. Pattern Recognition Letters, 171: 69–75, 2023. ISSN 0167-8655. doi: https://doi.org/10.1016/j.patrec. 2023.05.017. URL https://www.sciencedirect.com/science/article/pii/ S0167865523001447

  2. [2]

    S. H. Chan. Tilt-then-blur or blur-then-tilt? clarifying the atmospheric turbulence model. IEEE Signal Processing Letters , 29:1833–1837,

  3. [3]

    Chung, J

    H. Chung, J. Kim, S. Kim, and J. C. Ye. Parallel diffusion models of operator and image for blind inverse problems. 2023 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) , pages 6059–6069, 2022. URL https://api.semanticscholar.org/CorpusID: 253734166

  4. [4]

    D. L. Fried. Probability of getting a lucky short-exposure image through turbulence*. Journal of the Optical Society of America, page 1651, Dec

  5. [5]

    Gilles and N

    J. Gilles and N. B. Ferrante. Open turbulent image set (otis). Pattern Recognition Letters, page 38–41, Jan 2017. doi: 10.1016/j.patrec.2016. 12.020. URL https://doi.org/10.1016/j.patrec.2016.12.020

  6. [6]

    R. C. Hardie, J. D. Power, D. A. Lemaster, D. R. Droege, S. Gladysz, and S. R. Bose-Pillai. Simulation of anisoplanatic imaging through optical turbulence using numerical wave propagation with new val- idation analysis. Optical Engineering , 56, 2017. URL https://api. semanticscholar.org/CorpusID:59055918

  7. [7]

    Huang, X

    Z. Huang, X. Shi, C. Zhang, Q. Wang, K. C. Cheung, H. Qin, J. Dai, and H. Li. Flowformer: A transformer architecture for & nbsp;optical flow. In Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part XVII , page 668–685, Berlin, Heidelberg, 2022. Springer-Verlag. ISBN 978-3-031- 19789-5. doi: 10.100...

  8. [8]

    T.-W. Hui, X. Tang, and C. C. Loy. Liteflownet: A lightweight convo- lutional neural network for optical flow estimation. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8981– 8989, 2018. doi: 10.1109/CVPR.2018.00936

Show all 43 references
  1. [9]

    Ihrke, G

    I. Ihrke, G. Ziegler, A. Tevs, C. Theobalt, M. Magnor, and H.-P. Seidel. Eikonal rendering: efficient light transport in refractive objects. ACM Trans. Graph., 26(3):59–es, July 2007. ISSN 0730-0301. doi: 10.1145/ 1276377.1276451. URL https://doi.org/10.1145/1276377.1276451

  2. [10]

    Jaiswal, X

    A. Jaiswal, X. Zhang, S. H. Chan, and Z. Wang. Physics-driven turbulence image restoration with stochastic refinement. 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , pages 12136–12147, 2023. URL https://api.semanticscholar.org/ CorpusID:259991275

  3. [11]

    Jiang, V

    W. Jiang, V . Boominathan, and A. Veeraraghavan. Nert: Implicit neu- ral representations for unsupervised atmospheric turbulence mitiga- tion. In 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition Workshops (CVPRW), pages 4236–4243, 2023. doi: 10.1109/CVPRW5...

  4. [12]

    D. Jin, Y . Chen, Y . Lu, J. Chen, P. Wang, Z. Liu, S. Guo, and X. Bai. Neutralizing the impact of atmospheric turbulence on complex scene imaging via deep learning. Nature Machine Intelligence, 3:876 – 884,

  5. [13]

    E. C. C. Jr., A. W. Cooper, G. W. Rodeback, S. H. Kalmbach, and R. L. Armstead. Effects Of Turbulence On Imaging Through The Atmosphere. In R. Sepucha, editor, Optical Properties of the At- mosphere, volume 0142, pages 130 – 134. International Society for Optics and Photonics,...

  6. [14]

    A. N. Kolmogorov. The local structure of turbulence in incompressible viscous fluid for very large reynolds numbers.Proceedings of the Royal Society of London. Series A: Mathematical and Physical Sciences , 434:13 – 9, 1991. URL https://api.semanticscholar.org/CorpusID: 123612939

  7. [16]

    Liang, J

    J. Liang, J. Cao, Y . Fan, K. Zhang, R. Ranjan, Y . Li, R. Timofte, and L. Van Gool. Vrt: A video restoration transformer. IEEE Transac- tions on Image Processing , 33:2171–2182, 2024. doi: 10.1109/TIP. 2024.3372454

  8. [17]

    Liang, Y

    Y . Liang, Y . Fu, Y . Hu, W. Shao, J. Liu, and D. Zhang. Flow-anything: Learning real-world optical flow estimation from large-scale single- view images. IEEE Transactions on Pattern Analysis and Machine In- telligence, pages 1–18, 2025. doi: 10.1109/TPAMI.2025.3576851

  9. [18]

    López-Tapia, X

    S. López-Tapia, X. Wang, and A. K. Katsaggelos. Variational deep at- mospheric turbulence correction for video. In 2023 IEEE International Conference on Image Processing (ICIP), pages 3568–3572, 2023. doi: 10.1109/ICIP49359.2023.10222374

  10. [19]

    Z. Mao, N. Chimitt, and S. H. Chan. Accelerating atmospheric tur- bulence simulation via learned phase-to-space transform. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV) , pages 14739–14748, 2021. doi: 10.1109/ICCV48922.2021.01449

  11. [20]

    Z. Mao, A. Jaiswal, Z. Wang, and S. H. Chan. Single frame atmospheric turbulence mitigation: A benchmark study and a new physics-inspired transformer model. In S. Avidan, G. Brostow, M. Cissé, G. M. Farinella, and T. Hassner, editors,Computer Vision – ECCV 2022, pages 430–446,...

  12. [21]

    D. Qin, R. K. Saha, S. Jayasuriya, J. Ye, and N. Li. Unsupervised region-growing network for object segmentation in atmospheric tur- bulence. CoRR, abs/2311.03572, 2023. doi: 10.48550/ARXIV .2311. 03572. URL https://doi.org/10.48550/arXiv.2311.03572

  13. [22]

    M. C. Roggemann, B. M. Welsh, D. Montera, and T. A. Rhoadarmer. Method for simulating atmospheric turbulence phase effects for multi- ple time slices and anisoplanatic conditions. Appl. Opt., 34(20):4037– 4051, Jul 1995. doi: 10.1364/AO.34.004037. URL https://opg.optica. org/a...

  14. [23]

    R. K. Saha, E. Salcin, J. Kim, J. Smith, and S. Jayasuriya. Turbu- lence strength c n2 estimation from video using physics-based deep learning. Opt. Express , 30(22):40854–40870, Oct 2022. doi: 10. 1364/OE.469976. URL https://opg.optica.org/oe/abstract.cfm?URI= oe-30-22-40854

  15. [24]

    R. K. Saha, D. Qin, N. Li, J. Ye, and S. Jayasuriya. Turb-seg- res: A segment-then-restore pipeline for dynamic videos with atmo- spheric turbulence. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 25286–25296, 2024. URL https://api.semantics...

  16. [25]

    J. Schmidt. Numerical simulation of optical wave propagation: With examples in MATLAB. Schmidt, J.D., 01 2010. ISBN 9780819483270. doi: 10.1117/3.866274

  17. [26]

    Shimizu, S

    M. Shimizu, S. Yoshimura, M. Tanaka, and M. Okutomi. Super- resolution from image sequence under influence of hot-air optical tur- bulence. In 2008 IEEE Conference on Computer Vision and Pattern Recognition, pages 1–8, 2008. doi: 10.1109/CVPR.2008.4587525

  18. [27]

    D. Sun, X. Yang, M.-Y . Liu, and J. Kautz. Pwc-net: Cnns for optical flow using pyramid, warping, and cost volume. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8934– 8943, 2018. doi: 10.1109/CVPR.2018.00931

  19. [28]

    S. Sun, W. Ren, X. Gao, R. Wang, and X. Cao. Restoring images in & nbsp;adverse weather conditions via & nbsp;histogram transformer. In Computer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29 – October 4, 2024, Proceedings, Part XXII , page 111–129, B...

  20. [29]

    Teed and J

    Z. Teed and J. Deng. Raft: Recurrent all-pairs field transforms for opti- cal flow. In A. Vedaldi, H. Bischof, T. Brox, and J.-M. Frahm, editors, Computer Vision – ECCV 2020, pages 402–419, Cham, 2020. Springer International Publishing. ISBN 978-3-030-58536-5

  21. [30]

    Teed and J

    Z. Teed and J. Deng. Raft: Recurrent all-pairs field transforms for op- tical flow. In Computer Vision – ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Proceedings, Part II , page 402–419, Berlin, Heidelberg, 2020. Springer-Verlag. ISBN 978-3-030- 5853...

  22. [31]

    Z. Wang, X. Cun, J. Bao, W. Zhou, J. Liu, and H. Li. Uformer: A general u-shaped transformer for image restoration. In 2022 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) , pages 17662–17672, 2022. doi: 10.1109/CVPR52688.2022.01716

  23. [32]

    Y . Wu, K. Cheng, T. Cao, D. Zhao, and J. li. Semi-supervised correction model for turbulence-distorted images. Optics Express, 32, 05 2024. doi: 10.1364/OE.524559

  24. [33]

    X. Xu, P. Yang, H. Xian, and Y . Liu. Robust moving objects detection in long-distance imaging through turbulent medium. Infrared Physics & Technology, 100:87–98, 2019. ISSN 1350-4495. doi: https://doi.org/ 10.1016/j.infrared.2019.02.014. URL https://www.sciencedirect.com/ sci...

  25. [34]

    Yasarla and V

    R. Yasarla and V . M. Patel. Learning to restore images degraded by atmospheric turbulence using uncertainty. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1694–1698, 2021. doi: 10.1109/ICIP42928.2021.9506614

  26. [35]

    Yasarla and V

    R. Yasarla and V . M. Patel. Cnn-based restoration of a single face image degraded by atmospheric turbulence. IEEE Transactions on Biometrics, Behavior, and Identity Science , 4(2):222–233, 2022. doi: 10.1109/TBIOM.2022.3169697

  27. [36]

    Zamek and Y

    S. Zamek and Y . Yitzhaky. Turbulence strength estimation from an arbitrary set of atmospherically degraded images.Journal of the Optical Society of America. A, Optics, image science, and vision , 23 12:3106– 13, 2006. URL https://api.semanticscholar.org/CorpusID:924517

  28. [37]

    S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M. Yang. Restormer: Efficient transformer for high-resolution image restoration. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR) , pages 5718–5729, 2022. doi: 10.1109/CVPR52688. 2022.00564

  29. [38]

    Zhang, W

    K. Zhang, W. Zuo, Y . Chen, D. Meng, and L. Zhang. Beyond a gaussian denoiser: Residual learning of deep cnn for image denois- ing. IEEE Transactions on Image Processing , page 3142–3155, Jul

  30. [39]

    Zhang, Z

    X. Zhang, Z. Mao, N. Chimitt, and S. Chan. Imaging through the at- mosphere using turbulence mitigation transformer. IEEE Transactions on Computational Imaging, PP:1–14, 01 2024. doi: 10.1109/TCI.2024. 3354421

  31. [40]

    Zou and N

    Z. Zou and N. Anantrasirichai. Deturb: Atmospheric turbulence mitiga- tion with deformable 3d convolutions and 3d swin transformers. InPro- ceedings of the Asian Conference on Computer Vision (ACCV) , pages 904–921, December 2024

  32. [1978]

    URL http://dx.doi.org/10.1364/ josa.68.001651

    doi: 10.1364/josa.68.001651. URL http://dx.doi.org/10.1364/ josa.68.001651

  33. [2017]

    URL http://dx.doi.org/10.1109/ tip.2017.2662206

    doi: 10.1109/tip.2017.2662206. URL http://dx.doi.org/10.1109/ tip.2017.2662206

  34. [2021]

    URL https://api.semanticscholar.org/CorpusID:244578813

  35. [2022]

    doi: 10.1109/LSP.2022.3200551

Pith tools

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