Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Simulating Refractive Distortions and Weather-Induced Artifacts for Resource-Constrained Autonomous Perception

T0 review · 4 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 claims a procedural pipeline can turn ordinary dashcam frames from Southern African roads into paired clean/corrupted training data using classical optical and weather models, with three restoration baselines showing the…

desk verdict Standard distortion/weather models, packaged for African dashcams, with an internal benchmark and an unvalidated realism claim. read the letter →

arxiv 2507.05536 v1 pith:ZXTGGF5D submitted 2025-07-07 cs.CV cs.ETcs.LG

Moseli Mots'oehli , Feimei Chen , Hok Wai Chan , Itumeleng Tlali , Thulani Babeli , Kyungim Baek , Huaijin Chen This is my paper
classification cs.CVcs.ETcs.LG
keywords proceduralaugmentationrefractivedistortionsimulationweatherartifactdashcamperceptionimagerestorationlow-resourceautonomousdrivingAfricanroaddatasetsUVdisplacementfield
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

This paper tries to address the scarcity of autonomous-driving imagery from African roads without expensive sensor rigs, labeling campaigns, or high-budget simulation. It proposes a procedural pipeline that takes ordinary dashcam frames and applies classical optical and weather models — the extended Brown–Conrady lens model, Perlin and thin-plate-spline warps, divergence-free flow, Koschmieder fog with a Perlin extinction field, and a Gaussian lens-flare mask — to produce paired clean/corrupted images with the exact UV displacement maps or intensity maps used. Three restoration models, a residual U-Net, SegFormer-B1, and a denoising diffusion model, are trained on this synthetic data; the best results reported are 32.54 dB PSNR for weather artifacts and 23.24 dB for refractive distortion. If the pipeline's distortions really resemble what a dashcam would record on Southern African roads, the toolkit would let low-resource teams expand datasets and study local perception conditions at near-zero marginal cost.

What carries the argument

The load-bearing object is a dense UV displacement field, a per-pixel map $(x,y) \mapsto (x+\Delta x,\, y+\Delta y)$ that every refractive distortion is reduced to; the restoration models are trained to predict this flow and then warp the corrupted input through it. Radial distortion is generated by the extended Brown–Conrady polynomial, thermal shimmer by normalized Perlin fields with exponential covariance, bending by thin-plate spline interpolation between keypoint pairs, and turbulence by a divergence-free vector field $\mathbf{v}=(\partial\psi/\partial y,\, -\partial\psi/\partial x)$ built from a random stream function $\psi$. Weather artifacts use photometric maps instead of geometry: Koschmieder's law $I=I_{\text{clean}}\,t + A(1-t)$ with transmission $t=\exp(-k d)$, a linear depth gradient, and randomized extinction $k$ for uniform fog, a multi-scale Perlin $k$-map for heterogeneous fog, and a Gaussian mask for sun glare. The refractive training objective combines an L1 reconstruction loss with an L1 UV-flow loss, which is what ties accurate artifact modeling directly to the supervision signal.

What would settle it

Record the same routes with a real low-quality dashcam under genuine fog, sun glare, and heat haze, then train a restoration model on the synthetic pipeline and evaluate on those real frames; if it performs no better than a model trained on clean data, the central realism claim is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that a parameterized, purely procedural generator is enough to create supervised training data for removing refractive distortion and weather artifacts from dashcam frames. Starting from rectified frames, the refractive module warps every pixel through four mechanisms: the extended Brown–Conrady polynomial for barrel and cushion distortion, Perlin noise with exponential covariance for heat shimmer, a thin-plate spline for uneven-lens bending, and a divergence-free vector field derived from a scalar stream function for swirl-like turbulence. The weather module changes pixel intensities rather than geometry: uniform fog follows Koschmieder's law with transmission $t(x,y)=\exp(-k d(x,y))$ and a randomized extinction coefficient, heterogeneous fog varies the extinction through multi-scale Perlin noise, and lens flare is a Gaussian mask placed in the upper image region. On this synthetic data, the paper shows that a frozen-encoder ResUnet reaches 32.54 dB on deweathering and 23.24 dB on undistortion, the diffusion model performs comparably, and the smallest transformer underfits, with UV-map endpoint error confirming that flow prediction is the limiting factor for strong radial distortion. The claim is that this makes realistic training-data generation available wherever a cheap dashcam and modest compute exist.

Load-bearing premise

The whole benchmark rests on the assumption that these procedurally generated distortions and weather effects closely resemble what a real dashcam would record on Southern African roads, and the paper does not validate that resemblance against real distorted or foggy footage.

Editorial extensions

If this is right

  • The pipeline turns a commodity dashcam into a data generator: paired clean/corrupted frames plus ground-truth UV and intensity maps can be produced at scale without collection campaigns or physics-based rendering.
  • Weather restoration reaches high quality (best 32.54 dB PSNR), while refractive undistortion remains harder (best 23.24 dB), so lens-distortion correction is the task that needs the most future work.
  • The released toolkit, augmented splits, and three-model benchmark give low-resource teams a concrete starting point for comparing restoration approaches on African road imagery.
  • Because the UV map is supervised directly, any downstream model that needs geometric consistency can be trained with explicit flow targets rather than only image-level losses.

Reading between the lines

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

  • Not tested in the paper: if the realism premise holds, the same augmentations could benefit downstream perception tasks such as object detection and lane tracking, since restoration quality is a proxy for feature visibility; a detector benchmark on the augmented data would settle this.
  • The paper applies each warp independently per frame; smoothing the random UV fields in time to create temporally coherent video distortions would make the pipeline usable for video-based trackers and is a natural extension.
  • The realism assumption is the untested premise: a small validation set of real dashcam frames captured under actual fog, glare, and known cheap lenses would let the hand-picked parameter ranges be calibrated or refuted.
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 / 6 minor

Summary. The paper introduces a procedural augmentation pipeline that takes real dashcam frames recorded in four South African suburbs, rectifies them using calibrated camera parameters, and applies synthetic refractive distortions (extended Brown-Conrady radial distortion, Perlin warping, Thin-Plate Spline warping, and divergence-free warping) as well as weather-induced artifacts (uniform fog, heterogeneous fog, and lens flare). The authors train three restoration baselines (ResUnet, SegFormer-B1, and a conditional DDPM) on the resulting paired clean/corrupted data and report PSNR and endpoint-error metrics on a validation set generated by the same pipeline. The paper also releases the distortion toolkit and benchmark results as a resource for low-resource autonomous driving research in African contexts.

Significance. If the generated distortions are faithful to real optical and weather conditions on Southern African roads, the pipeline would be a useful, low-cost complement to dataset collection and high-budget simulation, and the released toolkit would support reproducible research in a region that is underserved by existing autonomous-driving benchmarks. The mathematical formulation of the distortions is standard and appears internally consistent, and the training setup is conventional. At the same time, the paper's central value proposition is the realism and utility of the augmentation, and that premise is not currently supported by external evidence; the benchmark as presented measures how well models invert the synthetic generator on its own distribution rather than how well the pipeline transfers to real perception conditions.

major comments (4)
  1. [Section 3.1 and Section 3.2] The central claim that the synthetic distortions 'closely resemble what an autonomous perception system would see on Southern African roads' (Section 3.1) is not validated. The parameter ranges for the Brown-Conrady coefficients, Perlin/TPS/divergence-free displacement amplitudes, extinction coefficient, depth model, flare radius, and flare intensity are all hand-picked, with no calibration against actual dashcam lenses, measured atmospheric conditions, or real fog/glare imagery from the four target suburbs. This is load-bearing because the paper's stated contribution is dataset expansion for real deployment, not merely a synthetic benchmark. Please provide at least one of: quantitative similarity metrics between synthetic and real distorted frames, calibration of parameter ranges to measured lens or weather data, perceptual or downstream-task evaluation on real distorted dashcam sequences, or a clearly stated reframing of the contributions as a purely synthetic benchmark without a realism claim.
  2. [Table 2 and Section 5] All reported validation metrics are computed on images generated by the same pipeline used to create the training data, from a single random 80/20 split with no error bars or seed variation. The results therefore establish that the models can approximately invert the synthetic generator, but they do not establish that the generator reproduces real distortions or that the restoration models will transfer to real dashcam footage. Please report metrics across multiple splits or seeds with standard deviations, and add at least one external evaluation: for example, testing on real distorted/foggy dashcam frames, testing on an independently generated synthetic set with different parameter distributions, or applying the trained models to a real downstream perception task such as detection or segmentation.
  3. [Section 4 and Table 2] The benchmark does not compare against any prior method, despite the related-work section citing relevant existing approaches such as Flare7K++, SimFIR, domain-adaptive fog detection, and other synthetic augmentation pipelines. Without at least one comparison baseline, it is unclear what the proposed pipeline adds over existing synthetic distortion/fog simulation or existing restoration methods. Please include comparisons against one or more prior augmentation schemes or restoration baselines, or justify why such comparisons are not feasible in the low-resource setting.
  4. [Section 3.1 and Section 3.3] Because the pipeline first rectifies the real frames using calibrated camera parameters and then applies synthetic random distortions, the 'clean' ground-truth image is not the true scene as captured by the original camera; it is a rectified and resampled version in which any real lens distortion present in the original footage has been removed and replaced by synthetic distortion. The paper should discuss the consequences of this for real-world deployment and provide evidence that the rectification/resampling step does not remove information needed by downstream perception systems (for example, by evaluating on the original unrectified frames or by comparing against a pipeline that skips rectification).
minor comments (6)
  1. [Equation (14)] The DDPM loss is written as L_DDPM = 0.8 ||\hat{\epsilon} - \epsilon||_1 + ||warp(x, \hat{y}) - GT_rgb||_2^2, but \hat{y} is not defined in the text; earlier the predicted UV flow is called \hat{y}_{uv}. Please make the notation consistent and clarify how the warped output is computed from the predicted flow.
  2. [Table 2] In the weather-artifact rows, the column Lk is labeled 'k-map/UV loss', but the deweathering objective defined in Section 4.1 contains only an image-space L1 term; it is unclear what Lk represents for the weather task. Please define this quantity or rename the column.
  3. [Section 4.2] The EPE formula is written ambiguously: it uses \|\hat{y}^{(i)}_{uv} - GT^{(i)}_{uv}\|_2 but does not state that this is the Euclidean norm of the two-component flow vector, nor are the units defined. Please spell out the formula and the units.
  4. [Section 3.3] The dataset description states that the four suburbs and the weather/time conditions are 'not all equally represented' but does not report the actual distribution. Please provide per-suburb and per-condition counts so that readers can assess class imbalance and potential biases.
  5. [Section 4.1] The statement 'We freeze the encoder weights and only allow decoder weight updates' is unclear for the DDPM model, which does not have a conventional encoder-decoder structure. Please specify which components of the DDPM are frozen and which are trained.
  6. [Figure 1] Figure 1 is difficult to read: the labels 'GSTOOLS', the flow arrows, and the module boundaries are small. Please enlarge the figure and add a more descriptive caption that maps each visual element to the corresponding section.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the synthetic pipeline and benchmark are self-consistent, and the unvalidated realism premise is an external-validity gap rather than a circular derivation.

full rationale

The paper's derivation chain is a data-generation pipeline followed by supervised restoration baselines. Each artifact is generated by explicit forward models: Brown-Conrady lens distortion (Eqs. 1-2), Perlin warping (Eq. 3), TPS warping (Eqs. 4-5), divergence-free warping (Eq. 6), Koschmieder uniform fog (Eqs. 7-8), heterogeneous fog (Eqs. 9-11), and Gaussian lens flare (Eqs. 12-13). The restoration targets, GT_uv and GT_rgb, are outputs of these same generators, and the losses in Section 4.1 supervise inversion of those generators. This is a self-consistent supervised benchmark, but it is not circular in the sense targeted here: no parameter is fitted to a subset and then reported as a prediction of that subset, and no claimed result is defined in terms of the quantity it purports to derive. The main weakness is the realism claim in Section 3.1, that the generated pairs 'closely resemble what an autonomous perception system would see on Southern African roads.' That claim is not validated against real distorted dashcam imagery or measured fog/glare statistics, so the benchmark's transferability is unsubstantiated. However, the generated data are not defined by, nor fitted to, real Southern African distortion measurements; the realism assertion is an unsupported premise rather than a conclusion that reduces to its own inputs. The only self-citation, reference [19], supports the general statement that developing-region datasets are scarce, and it is not load-bearing for any of the paper's equations or architectural choices. Accordingly, none of the enumerated circularity patterns is present; the appropriate finding is no significant circularity, score 0.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim of 'realistic' simulation rests entirely on hand-chosen parameters and standard synthetic models, none of which are validated against measurements from real low-cost cameras or real African weather. The free parameters are the sampling distributions for the distortions and weather effects; they are not fitted to data but they define the realism, and the paper does not justify them externally.

free parameters (7)
  • Perlin displacement amplitude α (refractive) = not specified; randomized
    Controls maximum pixel shift in Perlin heat-haze warp (Eq. 3). No calibration to real turbulence data.
  • Divergence-free warp scaling α = not specified; randomized
    Controls maximum displacement of the swirl-like warp in Section 3.1.4; chosen by hand.
  • Extinction coefficient k0 = 0.0375 (from visibility 100 m)
    Set to -ln(0.05)/100 in Section 3.2.1. A modeling choice for 'uniform fog'.
  • Maximum depth Dmax = 160.0 m
    Linear depth gradient parameter in Section 3.2.1; chosen by hand.
  • Perlin noise scales/weights (heterogeneous fog) = Table 1 (s_i, w_i)
    Hand-picked multi-scale weights for patchy fog in Section 3.2.2.
  • Lens flare radius ρ and intensity β = ρ ∈ [0.25, 0.35], β ∈ [0.55, 0.65]
    Randomized ranges in Section 3.2.3; not grounded in optic measurements.
  • Gaussian correlation length ℓ (Perlin) = not reported
    Underlies the exponential covariance in Section 3.1.2; never given a value, so reproducibility is limited.
assumptions (5)
  • standard math Extended Brown-Conrady lens distortion model (Eq. 1-2)
    Cited as prior work; used as ground-truth warp generator.
  • standard math Koschmieder's law for fog scattering (Eq. 7)
    Standard atmospheric scattering model; used to combine transmission map with clean image.
  • standard math Thin-plate spline interpolation (Eq. 4-5)
    Standard interpolation; cited from Bookstein.
  • standard math Divergence-free vector field from stream function (Eq. 6)
    Standard vector calculus construction.
  • domain assumption Perlin noise approximates atmospheric turbulence
    Assumed without physical validation that Perlin noise captures heat shimmer statistics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Simulating Refractive Distortions and Weather-Induced Artifacts for Resource-Constrained Autonomous Perception." pith.science (2026). https://pith.science/paper/ZXTGGF5D

@misc{pith2026250705536,
  author       = {Pith},
  title        = {Pith review of: Simulating Refractive Distortions and Weather-Induced Artifacts for Resource-Constrained Autonomous Perception},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZXTGGF5D}},
  note         = {Machine review of arXiv:2507.05536}
}
read the original abstract

The scarcity of autonomous vehicle datasets from developing regions, particularly across Africa's diverse urban, rural, and unpaved roads, remains a key obstacle to robust perception in low-resource settings. We present a procedural augmentation pipeline that enhances low-cost monocular dashcam footage with realistic refractive distortions and weather-induced artifacts tailored to challenging African driving scenarios. Our refractive module simulates optical effects from low-quality lenses and air turbulence, including lens distortion, Perlin noise, Thin-Plate Spline (TPS), and divergence-free (incompressible) warps. The weather module adds homogeneous fog, heterogeneous fog, and lens flare. To establish a benchmark, we provide baseline performance using three image restoration models. To support perception research in underrepresented African contexts, without costly data collection, labeling, or simulation, we release our distortion toolkit, augmented dataset splits, and benchmark results.

Figures

Figures reproduced from arXiv: 2507.05536 by the authors.

Figure 1
Figure 1. Overview of our two-stage restoration pipeline. The Synthetic Distortion Module applies random refractive distortions and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Visualization of refractive distortions for randomly selected samples: Columns (1) Undistorted frame, (2) Distorted frame, (3) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Example of the three simulated weather-induced artifacts (undistorted input, homogeneous fog, heterogeneous fog, and lens flare. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Refractive Distortion Training and Validation Curves. (a) Total Loss (train vs. val), showing steady convergence. (b) PSNR improvement over epochs for both train and val. (c) Endpoint Error (EPE) on UV flow decreases as training proceeds. 5 10 15 Epoch 0.1 0.2 0.3 0.4 …
Figure 5
Figure 5. Figure 5: Synthetic Weather-induced Artifact Training and Validation Curves. (a) Total loss (train vs. val). (b) PSNR improvement over epochs for both train and val. between the best and worst models seems to widen on the harder refractive undistortion task. In both tasks, a low…
Figure 6
Figure 6. Figure 6: Comparison of the undistorted input image, distorted image, and the best model’s undistortion of the input for all four types of [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Qualitative deweathering results for three weather-induced artifacts: uniform fog, heterogeneous fog, and lens flare. From left to [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

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. LLMs and Agentic AI in Insurance Decision-Making: Opportunities and Challenges For Africa

    cs.CE 2025-08 unverdicted novelty 3.0 of 10

    LLMs and agentic AI are presented as a transformative opportunity for African insurance, with a call for African-led, equitable AI strategies.

Reference graph

Works this paper leans on

30 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Principal Warps: Thin-Plate Splines and the Decomposition of Deformations

    Fred Bookstein. Principal Warps: Thin-Plate Splines and the Decomposition of Deformations. IEEE Transactions on Pattern Analysis and Machine Intelligence , 11(6):567–585,

    work page

  2. [2]

    The opencv library

    Gary Bradski. The opencv library. Dr. Dobb’s Journal: Soft- ware Tools for the Professional Programmer , 25(11):120– 123, 2000. 3

    work page 2000

  3. [3]

    Close-Range Camera Calibration

    Duane Brown. Close-Range Camera Calibration. Pho- togrammetric Engineering, 37(8):855–866, 1971. 1

    work page 1971

  4. [4]

    Celia Cook, Chris J. C. Reason, and Bruce C. Hewitson. Wet and dry spells within particularly wet and dry summers in the south african summer rainfall region. Climate Research, 26: 17–31, 2004. 5

    work page 2004

  5. [5]

    Flare7k++: Mixing syn- thetic and real datasets for nighttime flare removal and be- yond

    Yuekun Dai, Chongyi Li, Shangchen Zhou, Ruicheng Feng, Yihang Luo, and Chen Change Loy. Flare7k++: Mixing syn- thetic and real datasets for nighttime flare removal and be- yond. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(11):7041–7055, 2024. 2

    work page 2024

  6. [6]

    Divergence-Free Shape Interpolation and Correspondence

    Marvin Eisenberger, Zorah L ¨ahner, and Daniel Cremers. Divergence-free shape interpolation and correspondence. arXiv preprint arXiv:1806.10417, 2018. 3

    work page Pith review arXiv 2018

  7. [7]

    Simfir: A simple framework for fisheye image rectification with self-supervised representation learn- ing

    Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learn- ing. In 2023 IEEE/CVF International Conference on Com- puter Vision (ICCV), pages 12384–12393, 2023. 1

    work page 2023

  8. [8]

    Geostatistics for Natural Resources Eval- uation

    Pierre Goovaerts. Geostatistics for Natural Resources Eval- uation. Oxford University Press, 1997. 4

    work page 1997

Show all 30 references
  1. [9]

    Lilienthal

    Himanshu Gupta, Oleksandr Kotlyar, Henrik Andreasson, and Achim J. Lilienthal. Robust object detection in chal- lenging weather conditions. In 2024 IEEE/CVF Winter Con- ference on Applications of Computer Vision (WACV), pages 7508–7517, 2024. 2

    2024

  2. [10]

    Alarc´on

    Javier Gurrola-Ramos, Oscar Dalmau, and Teresa E. Alarc´on. A residual dense u-net neural network for image denoising. IEEE Access, 9:31742–31754, 2021. 2, 6

    2021

  3. [11]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. In Advances in Neural Informa- tion Processing Systems (NeurIPS), 2020. 2, 6

    2020

  4. [12]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. InPro- ceedings of the Tenth International Conference on Learning Representations (ICLR), 2022. 2

    2022

  5. [13]

    A novel vec- tor potential random flow generation method for synthesiz- ing divergence-free homogeneous isotropic turbulence with arbitrary spectra

    Chao Li, Lingwei Chen, Jinghan Wang, Wentong Zhang, Xi- angjie Wang, Zhuoran Wang, and Gang Hu. A novel vec- tor potential random flow generation method for synthesiz- ing divergence-free homogeneous isotropic turbulence with arbitrary spectra. Physics of Fluids, 36(3), 2024. 2

    2024

  6. [14]

    Domain adaptive object detection for au- tonomous driving under foggy weather

    Yifan Li, Yuwei Cui, Kui Zhao, Enze Xie, Jifeng Dai, and Dahua Lin. Domain adaptive object detection for au- tonomous driving under foggy weather. In Proceedings of the IEEE/CVF Winter Conference on Applications of Com- puter Vision (WACV), pages 1235–1244, 2023. 2

    2023

  7. [15]

    Diffusion restoration adapter for real-world image restoration

    Hanbang Liang, Zhen Wang, and Weihui Deng. Diffusion restoration adapter for real-world image restoration. arXiv preprint arXiv:2502.20679, 2025. 2

    2025 arXiv

  8. [16]

    Dafir: Distortion-aware representation learning for fisheye image rectification

    Zhaokang Liao, Wengang Zhou, and Houqiang Li. Dafir: Distortion-aware representation learning for fisheye image rectification. IEEE Transactions on Circuits and Systems for Video Technology, 34(5):3606–3618, 2024. 1

    2024

  9. [17]

    Comparison of sim- ulation methods for glare risk assessment with roller shades

    Sichen Lu and Athanasios Tzempelikos. Comparison of sim- ulation methods for glare risk assessment with roller shades. Buildings, 14(6):1773, 2024. 5

    2024

  10. [18]

    Evenformer: Dynamic even trans- former for real-world image restoration

    Xin Lu, Yuanfei Bao, Jiarong Yang, Anya Hu, Jie Xiao, Kunyu Wang, Dong Li, Senyan Xu, Kean Liu, Xueyang Fu, and Zheng-Jun Zha. Evenformer: Dynamic even trans- former for real-world image restoration. InCVPR Workshops (NTIRE), pages 22–38, 2025. 2

    2025

  11. [19]

    Public parking spot detection and geo-localization using transfer learning

    Moseli Mots’oehli and Yao Chao Yang. Public parking spot detection and geo-localization using transfer learning. In Proceedings of the Southern African Conference on Artifi- cial Intelligence Research (SACAIR), 2022. 1

    2022

  12. [20]

    GSTools v1.3: a toolbox for geostatistical mod- elling in Python

    Sebastian M ¨uller, Lennart Sch ¨uler, Andreas Zech, and Flo- rian Heße. GSTools v1.3: a toolbox for geostatistical mod- elling in Python. Geoscientific Model Development , 15: 3161–3182, 2022. 3

    2022

  13. [21]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In 2023 IEEE/CVF International Confer- ence on Computer Vision (ICCV) , pages 4172–4182, 2023. 2

    2023

  14. [22]

    An Image Synthesizer

    Ken Perlin. An Image Synthesizer. In Proceedings of the 12th Annual Conference on Computer Graphics and Inter- active Techniques (SIGGRAPH), pages 287–296, 1985. 1, 3

    1985

  15. [23]

    U- Net: Convolutional Networks for Biomedical Image Seg- mentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U- Net: Convolutional Networks for Biomedical Image Seg- mentation. In Medical Image Computing and Computer- Assisted Intervention (MICCAI), pages 234–241, 2015. 2

    2015

  16. [24]

    Fleet, and Mohammad Norouzi

    Chitwan Saharia, Jonathan Ho, William Chan, Tim Sali- mans, David J. Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In Proceedings of the ACM International Conference on Machine Learning (ICML), 2022. 2

    2022

  17. [25]

    The Effects of the Turbulent Atmosphere on Wave Propagation

    Valerian Tatarskii. The Effects of the Turbulent Atmosphere on Wave Propagation. National Technical Information Ser- vice, U.S. Dept. of Commerce, 1971. 1

    1971

  18. [26]

    Alvarez, and Ping Luo

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transform- ers. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 2, 6

    2021

  19. [27]

    Diff-unet: A diffusion embedded network for volumet- 9 ric segmentation

    Zhaohu Xing, Liang Wan, Huazhu Fu, Guang Yang, and Lei Zhu. Diff-unet: A diffusion embedded network for volumet- 9 ric segmentation. arXiv preprint arXiv:2303.10326 , 2023. 2

    2023 arXiv

  20. [28]

    Real-time rendering of heterogeneous fog based on the graphics hardware acceleration

    Dorota Zdrojewska. Real-time rendering of heterogeneous fog based on the graphics hardware acceleration. In Central European Seminar on Computer Graphics (CESCG) , 2004. 5

    2004

  21. [29]

    Enhancing object detection in ad- verse weather conditions through entropy and guided multi- modal fusion

    Zhenrong Zhang, Haoyan Gong, Yuzheng Feng, Zixuan Chu, and Hongbin Liu. Enhancing object detection in ad- verse weather conditions through entropy and guided multi- modal fusion. In Proceedings of the 17th Asian Conference on Computer Vision (ACCV), page 22–38, 2024. 2

    2024

  22. [30]

    Unet++: Redesigning skip connections to exploit multiscale features in image segmen- tation

    Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang. Unet++: Redesigning skip connections to exploit multiscale features in image segmen- tation. IEEE Transactions on Medical Imaging, 2019. 2 10

    2019

Pith tools

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