pith. sign in

arxiv: 2606.29093 · v1 · pith:SYHB2XF6new · submitted 2026-06-27 · 💻 cs.CV · physics.optics

From Fog Chamber to Aircraft Window: Pixel-Registered Imaging and Synthetic Fine-Tuning Enable Cross-Domain Defogging

Alexander Ingold , Sabina D. Menon , Manya Yellepeddy , Alec Ikei , John D. Hodges , Jordan Baker , Syed N. Qadri , Rajesh Menon This is my paper

Pith reviewed 2026-06-30 09:25 UTC · model grok-4.3

classification 💻 cs.CV physics.optics
keywords defoggingimage restorationdomain transfersynthetic datapaired supervisioncross-domain generalizationfog chamber
0
0 comments X

The pith

A defogging network pretrained on lab fog pairs and fine-tuned with randomized synthetic fog generalizes to aircraft-window video from an unseen camera with no target training.

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

The paper establishes that a restoration network first trained on exactly registered foggy and clear image pairs captured through a fixed scattering path in a lab chamber can be adapted by overlaying randomized synthetic fog on clear outdoor scenes. This two-stage process produces outputs that remain stable and improve quality metrics when applied to free-flowing real fog and to video recorded through an aircraft cabin window using an iPhone. The registration step supplies pixel-exact supervision that supports direct L1 training and a paired quality predictor, while the randomization step covers variations in fog density, airlight, and noise. A reader would care because the result removes the need to collect new paired data for each new sensor or optical path when building practical defogging systems.

Core claim

Pretraining on 5,495 pixel-aligned foggy/clear pairs obtained by imaging a flat-panel display through a 114 mm artificial-fog enclosure, followed by fine-tuning on clear outdoor scenes with on-the-fly randomized synthetic fog, produces a model that transfers to a graded sequence of out-of-distribution settings including chamber-free fog and iPhone video through an aircraft cabin window, with NIQE falling from 6.22 to 4.97 on the latter and temporally stable output across motion sequences.

What carries the argument

Pixel-registered laboratory imaging that supplies paired Laplacian-ratio quality prediction and exact L1 supervision, combined with domain-randomized synthetic fog fine-tuning that spans strength, spatial variation, airlight, and noise.

If this is right

  • On a held-out 552-image split the top backbone reaches 24.33 dB PSNR and 0.7912 SSIM while a compact alternative stays within 1.29 dB at 3 percent of the parameter count.
  • A ResNet-50 classifier verifies that restored images retain semantic content rather than only low-level structure.
  • The same pipeline under paired supervision reaches 20.71 dB / 0.683 SSIM on a non-overlapping O-HAZE/NH-HAZE split.
  • The paired Laplacian ratio predicts restoration quality with Spearman rho of 0.632, outperforming single-image proxies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same registration-plus-randomization pattern could be tested on other single-image degradations such as rain streaks or low-light noise.
  • If the synthetic-fog distribution is expanded further, the approach might support direct deployment on additional unseen sensors without any fine-tuning step.
  • The exact pixel alignment also enables controlled ablation of the scattering-path length to isolate which optical parameters matter most for transfer.

Load-bearing premise

Randomized synthetic fog overlaid on clear outdoor scenes sufficiently matches the scattering, sensor response, and statistics of the aircraft-window optical path.

What would settle it

Failure of the fine-tuned model to reduce NIQE or to maintain temporal stability on the aircraft-window video sequences would falsify the zero-shot transfer claim.

Figures

Figures reproduced from arXiv: 2606.29093 by Alec Ikei, Alexander Ingold, John D. Hodges, Jordan Baker, Manya Yellepeddy, Rajesh Menon, Sabina D. Menon, Syed N. Qadri.

Figure 1
Figure 1. Figure 1: Controlled fog-chamber training, benchmark results, and generalization to [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Testing our trained model on videos captured using an iPhone camera through [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Task-specific public-haze and nighttime-haze checks. (a) O-HAZE, (b) NH [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Defogging benchmark across established models on the fog-chamber dataset. (a) [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
read the original abstract

A deep defogging pipeline pretrained on controlled laboratory fog and fine-tuned with domain-randomized synthetic fog applied to clear outdoor scenes generalizes across a graded sequence of out-of-distribution settings with no target-domain training, from chamber-free free-flowing fog to iPhone video recorded through an aircraft cabin window in flight, an entirely unseen sensor, scene, and optical path. This directly addresses an open transfer limitation reported for real-world binocular defogging. Two design choices support the transfer. First, a single-camera fog imager photographs a flat-panel display through an artificial-fog enclosure with a fixed 114~mm scattering path, producing 5{,}495 pixel-aligned foggy/clear pairs. Exact registration permits a paired Laplacian ratio that predicts per-image restoration quality far better than single-image proxies (Spearman $\rho = 0.632$ versus $0.399$) and supports pixel-exact $L_1$ reconstruction training that avoids adversarial hallucination. Second, the fog-chamber checkpoint is fine-tuned on Mapillary Vistas crops overlaid with on-the-fly randomized synthetic fog spanning a broad range of strengths, spatial variations, airlights, and noise conditions. On a 552-image held-out split, a uniform comparison of 30 restoration backbones places NAFNet at the top (24.33~dB~/~0.7912~SSIM), with a compact alternative within 1.29~dB at 3\% of the parameter count, and a ResNet-50 classifier confirms that the restoration preserves semantic content rather than only pixel-level structure. On unpaired aircraft-window video, NIQE decreases from a mean of 6.22 to 4.97 after fine-tuning, with temporally stable output across full-motion sequences. The same backbone, under paired supervision, also reaches 20.71~dB~/~0.683~SSIM on a non-overlapping O-HAZE/NH-HAZE split (a transferability check rather than a competitive ranking).

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a defogging pipeline pretrained on 5,495 pixel-registered foggy/clear image pairs captured in a laboratory fog chamber (single-camera imager viewing a flat-panel display through a fixed 114 mm scattering path) and subsequently fine-tuned on Mapillary Vistas crops overlaid with on-the-fly domain-randomized synthetic fog. It reports quantitative results on a 552-image held-out split (NAFNet reaching 24.33 dB PSNR / 0.7912 SSIM), semantic preservation via a ResNet-50 classifier, competitive transfer performance on a non-overlapping O-HAZE/NH-HAZE split (20.71 dB / 0.683 SSIM), and a NIQE reduction from 6.22 to 4.97 on unpaired real aircraft-cabin-window video, claiming zero-shot generalization across graded out-of-distribution settings without any target-domain training.

Significance. If the reported generalization holds, the work would meaningfully advance practical cross-domain defogging by combining controlled paired laboratory data with synthetic augmentation, directly addressing the open transfer limitation noted for real-world binocular defogging. Strengths include the exact pixel registration enabling paired L1 training and the Laplacian-ratio quality predictor (Spearman ρ = 0.632), which are concrete methodological contributions. The quantitative gains on the held-out split and the demonstration of temporal stability on real video further support potential utility in computer-vision restoration tasks.

major comments (2)
  1. [Abstract (aircraft-window video evaluation)] The central zero-shot transfer claim to the aircraft cabin window (unseen sensor, scene geometry, and optical path) rests on NIQE improvement (6.22 → 4.97) and temporal stability on unpaired video. No domain-similarity metric, fog-parameter estimation on real frames, or ablation removing the randomization ranges is reported to verify that the synthetic fog distribution on Mapillary Vistas approximates the target optical scattering, window reflections, polarization, or iPhone sensor response. This is load-bearing for the generalization assertion.
  2. [Abstract (O-HAZE/NH-HAZE evaluation)] The O-HAZE/NH-HAZE result is described as a transferability check rather than a competitive ranking, yet the manuscript provides no details on overlap or distributional distance between the Mapillary-based fine-tuning distribution and the O-HAZE scenes, making it difficult to quantify the degree of out-of-distribution generalization achieved.
minor comments (2)
  1. The randomization ranges for synthetic fog parameters (strength, spatial variation, airlight, noise) are described only qualitatively; a table or explicit interval list would improve reproducibility.
  2. A side-by-side visual comparison of synthetic fog overlays versus example aircraft-window frames would help readers assess domain similarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and agree that targeted additions will strengthen the generalization claims.

read point-by-point responses

  1. Referee: [Abstract (aircraft-window video evaluation)] The central zero-shot transfer claim to the aircraft cabin window (unseen sensor, scene geometry, and optical path) rests on NIQE improvement (6.22 → 4.97) and temporal stability on unpaired video. No domain-similarity metric, fog-parameter estimation on real frames, or ablation removing the randomization ranges is reported to verify that the synthetic fog distribution on Mapillary Vistas approximates the target optical scattering, window reflections, polarization, or iPhone sensor response. This is load-bearing for the generalization assertion.

    Authors: We acknowledge the absence of explicit domain-similarity metrics or per-frame fog-parameter estimation. The on-the-fly randomization spans wide ranges of scattering strength, spatial variation, airlight, and noise precisely to promote robustness across unseen optical paths and sensors; the laboratory paired data supplies the core pixel-exact supervision that enables this transfer. The NIQE drop and temporal stability on the iPhone aircraft video constitute the empirical support for the zero-shot claim. In revision we will add (i) a short discussion of the randomization hyper-parameter ranges relative to typical real scattering and (ii) an ablation that narrows those ranges to test sensitivity. We cannot retroactively compute polarization or iPhone-specific response functions without additional hardware data. revision: partial

  2. Referee: [Abstract (O-HAZE/NH-HAZE evaluation)] The O-HAZE/NH-HAZE result is described as a transferability check rather than a competitive ranking, yet the manuscript provides no details on overlap or distributional distance between the Mapillary-based fine-tuning distribution and the O-HAZE scenes, making it difficult to quantify the degree of out-of-distribution generalization achieved.

    Authors: The manuscript already labels the O-HAZE/NH-HAZE numbers as a transferability check on a non-overlapping split rather than a benchmark ranking. Mapillary Vistas supplies diverse outdoor urban and natural scenes; O-HAZE/NH-HAZE use different cameras, haze densities, and scene geometries. We will insert a brief paragraph in the revision that notes the absence of shared scenes and the differing capture conditions, thereby clarifying the intended degree of distributional shift without claiming a quantitative distance metric. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical results on independent held-out and unpaired real data

full rationale

The paper's central claims rely on a two-stage training process (lab-paired pretraining followed by synthetic fine-tuning on Mapillary Vistas) evaluated on a held-out lab split and unpaired aircraft-window video using the external NIQE metric. No equations, fitted parameters, or self-citations are shown to reduce the reported generalization metrics or transfer performance to quantities defined by the training inputs themselves. The Laplacian ratio is presented only as a quality predictor on the paired lab data and does not enter the main cross-domain results. This is a standard empirical pipeline whose outputs are not forced by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that synthetic fog randomization can bridge the gap between lab and aircraft optical paths, plus standard assumptions in neural image restoration training.

free parameters (1)
  • synthetic fog randomization ranges
    Strengths, spatial variations, airlights, and noise levels chosen to span real conditions; specific ranges are design choices.
axioms (1)
  • domain assumption Synthetic fog overlay on clear images approximates real atmospheric scattering for the purpose of domain adaptation.
    Invoked when fine-tuning the lab checkpoint on Mapillary Vistas crops.

pith-pipeline@v0.9.1-grok · 5949 in / 1344 out tokens · 40756 ms · 2026-06-30T09:25:15.731803+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references

  1. [1]

    Theorie der horizontalen Sichtweite,

    H. Koschmieder, “Theorie der horizontalen Sichtweite,” Beiträge zur Physik der freien Atmosphäre12, 33–53 (1924)

    1924

  2. [2]

    Vision and the atmosphere,

    S. G. Narasimhan and S. K. Nayar, “Vision and the atmosphere,” Int. J. Comput. Vis.48, 233–254 (2002)

    2002

  3. [3]

    Single image haze removal using dark channel prior,

    K. He, J. Sun, and X. Tang, “Single image haze removal using dark channel prior,” IEEE Trans. on Pattern Anal. Mach. Intell.33, 2341–2353 (2011)

    2011

  4. [4]

    Single image dehazing using color attenuation prior,

    Q. Zhu, J. Mai, and L. Shao, “Single image dehazing using color attenuation prior,” inBritish Machine Vision Conference,(2014)

    2014

  5. [5]

    DehazeNet: An end-to-end system for single image haze removal,

    B. Cai, X. Xu, K. Jia,et al., “DehazeNet: An end-to-end system for single image haze removal,” IEEE Trans. on Image Process.25, 5187–5198 (2016)

    2016

  6. [6]

    AOD-Net: All-in-one dehazing network,

    B. Li, X. Peng, Z. Wang,et al., “AOD-Net: All-in-one dehazing network,” inProceedings of the IEEE International Conference on Computer Vision,(2017), pp. 4770–4778

    2017

  7. [7]

    GridDehazeNet: Attention-based multi-scale network for image dehazing,

    X. Liu, Y. Ma, Z. Shi, and J. Chen, “GridDehazeNet: Attention-based multi-scale network for image dehazing,” in Proceedings of the IEEE/CVF International Conference on Computer Vision,(2019), pp. 7314–7323

    2019

  8. [8]

    FFA-Net: Featurefusionattentionnetworkforsingleimagedehazing,

    X.Qin,Z.Wang,Y.Bai,etal.,“FFA-Net: Featurefusionattentionnetworkforsingleimagedehazing,”inProceedings of the AAAI Conference on Artificial Intelligence,vol. 34 (2020), pp. 11908–11915

    2020

  9. [9]

    Vision transformers for single image dehazing,

    Y. Song, Z. He, H. Qian, and X. Du, “Vision transformers for single image dehazing,” IEEE Trans. on Image Process. 32, 1927–1941 (2023)

    1927

  10. [10]

    Simple baselines for image restoration,

    L. Chen, X. Chu, X. Zhang, and J. Sun, “Simple baselines for image restoration,” inEuropean Conference on Computer Vision,(2022), pp. 17–33

    2022

  11. [11]

    A comprehensive survey on image dehazing based on deep learning,

    J. Gui, X. Cong, Y. Cao,et al., “A comprehensive survey on image dehazing based on deep learning,” inProceedings of the Thirtieth International Joint Conference on Artificial Intelligence,(2021), pp. 4426–4433

    2021

  12. [12]

    TheSYNTHIAdataset: Alargecollectionofsyntheticimagesforsemantic segmentation of urban scenes,

    G.Ros,L.Sellart,J.Materzynska,etal.,“TheSYNTHIAdataset: Alargecollectionofsyntheticimagesforsemantic segmentation of urban scenes,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2016)

    2016

  13. [13]

    Semantic foggy scene understanding with synthetic data,

    C. Sakaridis, D. Dai, and L. Van Gool, “Semantic foggy scene understanding with synthetic data,” Int. J. Comput. Vis.126, 973–992 (2018)

    2018

  14. [14]

    Model adaptation with synthetic and real data for semantic dense foggy scene understanding,

    C. Sakaridis, D. Dai, S. Hecker, and L. Van Gool, “Model adaptation with synthetic and real data for semantic dense foggy scene understanding,” inProceedings of the European Conference on Computer Vision,(2018), pp. 687–704

    2018

  15. [15]

    Image-to-image translation with conditional adversarial networks,

    P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros, “Image-to-image translation with conditional adversarial networks,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,(2017), pp. 1125–1134

    2017

  16. [16]

    Unpaired image-to-image translation using cycle-consistent adversarial networks,

    J.-Y. Zhu, T. Park, P. Isola, and A. A. Efros, “Unpaired image-to-image translation using cycle-consistent adversarial networks,” inProceedings of the IEEE International Conference on Computer Vision,(2017), pp. 2223–2232

    2017

  17. [17]

    O-HAZE: A dehazing benchmark with real hazy and haze-free outdoor images,

    C. O. Ancuti, C. Ancuti, C. De Vleeschouwer, and P. Bekaert, “O-HAZE: A dehazing benchmark with real hazy and haze-free outdoor images,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops,(2018)

    2018

  18. [18]

    NH-HAZE: An image dehazing benchmark with non-homogeneous hazy and haze-free images,

    C. O. Ancuti, C. Ancuti, and R. Timofte, “NH-HAZE: An image dehazing benchmark with non-homogeneous hazy and haze-free images,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops,(2020)

    2020

  19. [19]

    NTIRE 2021 nonhomogeneous dehazing challenge report,

    C. O. Ancuti, C. Ancuti, F.-A. Vasluianu,et al., “NTIRE 2021 nonhomogeneous dehazing challenge report,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops,(2021), pp. 627–646

    2021

  20. [20]

    NTIRE 2026 nighttime image dehazing challenge,

    NTIRE 2026 Nighttime Image Dehazing Challenge organizers, “NTIRE 2026 nighttime image dehazing challenge,” Codabench competition, https://www.codabench.org/competitions/12855/ (2026). Accessed June 5, 2026

    2026

  21. [21]

    Prompt-based test-time real image dehazing: A novel pipeline,

    Z. Chen, Z. He, Z. Lu,et al., “Prompt-based test-time real image dehazing: A novel pipeline,” inProceedings of the European Conference on Computer Vision,(2024)

    2024

  22. [22]

    Tokenize image patches: Global context fusion for effective haze removal in large images,

    J. Chen, X. Yan, Q. Xu, and K. Li, “Tokenize image patches: Global context fusion for effective haze removal in large images,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,(2025), pp. 2258–2268

    2025

  23. [23]

    Image-to-image machine translation enables computational defogging in real-world images,

    A. Pollak and R. Menon, “Image-to-image machine translation enables computational defogging in real-world images,” Opt. Express32, 33852–33860 (2024)

    2024

  24. [24]

    SpecAT: Spatial-spectral cumulative-attention transformer for high-resolution hyperspectral image reconstruction,

    Z. Yao, S. Liu, X. Yuan, and L. Fang, “SpecAT: Spatial-spectral cumulative-attention transformer for high-resolution hyperspectral image reconstruction,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR),(2024), pp. 25368–25377

    2024

  25. [25]

    130K images (512×512) universal image embeddings,

    R. H. Singh, “130K images (512×512) universal image embeddings,” Kaggle dataset, https://www.kaggle.com/ datasets/rhtsingh/130k-images-512x512-universal-image-embeddings (2026). Accessed June 5, 2026

    2026

  26. [26]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” inInternational Conference on Learning Representations (ICLR),(2015)

    2015

  27. [27]

    Image quality assessment: From error visibility to structural similarity,

    Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: From error visibility to structural similarity,” IEEE Trans. on Image Process.13, 600–612 (2004)

    2004

  28. [28]

    The unreasonable effectiveness of deep features as a perceptual metric,

    R. Zhang, P. Isola, A. A. Efros,et al., “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR),(2018), pp. 586–595

    2018

  29. [29]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition,(2016), pp. 770–778

    2016

  30. [30]

    TheMapillaryVistasdatasetforsemanticunderstanding of street scenes,

    G.Neuhold,T.Ollmann,S.RotaBulo,andP.Kontschieder,“TheMapillaryVistasdatasetforsemanticunderstanding of street scenes,” inProceedings of the IEEE International Conference on Computer Vision,(2017), pp. 5000–5009

    2017

  31. [31]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” inInternational Conference on Learning Representations (ICLR),(2019)

    2019

  32. [32]

    Ntire 2026 nighttime image dehazing challenge report,

    R. P. Ancuti, A. Brateanu, R. Balmez,et al., “Ntire 2026 nighttime image dehazing challenge report,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,(2026), pp. 1608–1625

    2026

  33. [33]

    Nt-haze: Abenchmarkdatasetforrealisticnight-timeimagedehazing,

    R.P.Ancuti,A.Brateanu,C.O.Ancuti,etal.,“Nt-haze: Abenchmarkdatasetforrealisticnight-timeimagedehazing,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition,(2026), pp. 1598–1607

    2026

  34. [34]

    Making a completely blind image quality analyzer,

    A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a completely blind image quality analyzer,” IEEE Signal Process. Lett.20, 209–212 (2013)

    2013