Pith. sign in

REVIEW 4 major objections 5 minor 118 references

Examining the Impact of Optical Aberrations to Image Classification and Object Detection Models

T0 review · 4 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Real optical blur re-ranks vision models differently than the standard defocus-blur test, two new benchmarks show.

desk verdict A genuinely useful optical-blur robustness benchmark, but the headline claim that kernel shape drives the weak correlation with disk blur is not isolated from chromatic aberration. read the letter →

arxiv 2504.18510 v1 pith:MMNWDXDV submitted 2025-04-25 cs.CV

Patrick Müller , Alexander Braun , Margret Keuper This is my paper · ORCID
classification cs.CV
keywords opticalaberrationsrobustnessbenchmarkZernikepolynomialspointspreadfunctionimageclassificationobjectdetectiondataaugmentationlensblur
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 argues that the standard way of testing vision models against blur—convolving with simple disk-shaped defocus kernels—does not rank models the way realistic optical blur does. It introduces two benchmark families: OpticsBench, which varies single Zernike aberrations (coma, astigmatism, defocus-and-spherical, trefoil), and LensCorruptions, which uses ray-traced point spread functions from 100 real lens prescriptions. Across ImageNet classification and MSCOCO/NuImages detection, model performance varies with the aberration type, and the rank correlation between the disk-blur baseline and the optical corruptions is weak (Kendall tau around 0.2–0.4, with coma not statistically significant). The paper also shows that training with an aberration-aware augmentation (OpticsAugment) closes much of the gap on OpticsBench and transfers some robustness to standard 2D corruptions. A fair reader should take away that realistic blur kernels belong in robustness evaluation.

What carries the argument

The engine is the point spread function (PSF), derived from the wavefront aberration $W_\lambda$ expanded into Zernike polynomials for single aberrations, or obtained by tracing rays through a lens prescription following the Huygens principle for real lenses. A critical modeling step matches each lens's PSF to a virtual pixel size using an MTF20 criterion: the sensor's Nyquist frequency is set so the lens retains 20% contrast at that frequency, fixing the otherwise free pixel-size parameter and yielding pixel sizes between 1 and 20 $\mu$m. The PSFs are then convolved with images; for LensCorruptions each lens supplies five field positions and three azimuth orientations, giving 500 corrupted datasets.

What would settle it

Take a set of real cameras with measured PSFs (for example from a lens bench), generate the same corruptions with those measured PSFs, and rank a dozen vision models on them; if the resulting ranking matches the ranking on the disk-defocus baseline with high Kendall tau, the paper's central claim about disk kernels being weak proxies would be contradicted. A cheaper check: replace the MTF20 rule with MTF50 or a fixed pixel pitch and test whether the LensCorruptions rankings change materially.

Watch

Extended reading notes

Core claim

The central discovery is that the shape of the blur kernel matters for model robustness, not just its size. Using Zernike polynomials to encode wavefront aberrations and Huygens ray tracing on real lens prescriptions, the authors generate point spread functions for coma, astigmatism, defocus-and-spherical, trefoil, and 100 real lenses. Across more than 65 classifiers and 8 detectors, they find that a model's rank on the disk-defocus baseline from Hendrycks et al. is only weakly predictive of its rank on these optical corruptions: Kendall tau values cluster around 0.2–0.4, and the coma ranking does not correlate significantly with the baseline (p = 0.5). The authors conclude that the disk-shaped kernel is a weak proxy for realistic optical degradation and that optical kernels should be part of robustness evaluation.

Load-bearing premise

The whole comparison assumes that PSFs computed from nominal lens prescriptions, with the virtual pixel size fixed by the MTF20 criterion, faithfully represent what a real camera of that lens would produce; if real lenses deviate from their nominal designs beyond manufacturing tolerances, or if the MTF20 rule does not reflect how sensors are actually matched, then the LensCorruptions severities and the comparison to disk blur are off.

Editorial extensions

If this is right

  • Robustness rankings from defocus-blur benchmarks do not transfer to optical aberrations, so benchmarks that use only disk-shaped kernels can mislead model selection for cameras in safety-critical systems.
  • Models trained with OpticsAugment gain on average 18% accuracy on OpticsBench (ImageNet-100) and improve robustness to standard 2D common corruptions by about 5.3 points, suggesting that aberration-aware augmentation is a practical remedy.
  • Lens quality as measured by MTF50 correlates strongly with model accuracy (Pearson $\rho \approx 0.86$ for classification and $\rho \approx 0.89$ for detection), meaning optical design quality can directly predict vision performance.
  • Accuracy and mAP drop with distance from the lens center, so field-dependent optical effects are measurable and should be included in camera-system evaluation.

Reading between the lines

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

  • If the weak rank correlation holds across a wider model zoo, then any robustness evaluation that reports a single 'blur' number is likely averaging over incompatible failure modes; splitting blur by aberration type would give model developers more actionable information.
  • The MTF20-based pixel-size rule implies a testable prediction: real cameras whose pixel pitch deviates from MTF20 matching should show even larger discrepancies from OpticsBench severities than the simulated ones.
  • LensCorruptions could be extended into a training-time regularizer by sampling lens PSFs directly, potentially covering the gap between simple disk blur and full optical models without the need for a separate augmentation step.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 two new blur-corruption benchmarks for vision models: OpticsBench, built from Zernike-mode PSFs for primary aberrations (astigmatism, coma, defocus with spherical, trefoil), and LensCorruptions, built from Huygens-traced PSFs of 100 curated real lens prescriptions. The authors evaluate 72 ImageNet classifiers, 8 MSCOCO detectors (plus NuImages), report TIDE error decompositions, and compare model rankings on their corruptions against the disk-shaped defocus baseline of Hendrycks et al. They report weak Kendall rank correlations and argue that disk-shaped kernels are weak proxies for optical blur. They also propose an augmentation method, OpticsAugment, trained on OpticsBench kernels, and report gains on OpticsBench and 2D common corruptions. The central claims are that realistic optical kernel shapes matter for robustness evaluation and that optical kernels are a useful training augmentation.

Significance. If the central claim holds, the paper makes a useful benchmark contribution: it provides two reusable corruption suites, a large lens-derived kernel database, a validation of a 1000-image COCO subset against the full set (0.2% mAP difference), and an unusually broad evaluation across architectures, including TIDE error decompositions. The LensCorruptions curation from 718 lenses and the explicit code release are also valuable. However, the load-bearing comparison that supports the 'shape matters' conclusion is confounded: OpticsBench kernels differ from the disk baseline both in PSF geometry and in chromatic structure, so the weak rank correlations do not identify shape as the cause. The realism of LensCorruptions also rests on an untested pixel-size and nominal-lens assumption. These issues are fixable with additional analyses, but they are central to the paper's message, so the manuscript needs revision before the claims are supported.

major comments (4)
  1. [§VI-A, Fig. 5, Table III; Suppl. E-A, Table XVII] The central claim that disk-shaped blur kernels are weak proxies because of 'different blur kernel shapes' is not identified by the reported experiment. The OpticsBench corruptions are generated by Eq. 5 per wavelength with a chromatic baseline (Suppl. E-A, Table XVII gives different Zernike coefficients for R, G, B), so each OpticsBench kernel is an RGB kernel with color fringes, while the Hendrycks baseline applies the same grayscale disk to all channels. The comparison therefore varies two factors at once: kernel geometry and chromatic/color-channel structure. The weak Kendall tau values (around 0.2–0.4, with coma non-significant at p=0.5) could be driven entirely by chromatic aberration. Please add an ablation that removes the color structure, e.g. grayscale versions of the OpticsBench kernels or disk kernels carrying the same per-channel chromatic variation, so that the shape hypothesis is tested in isolation.
  2. [§III-B and §V-A] The realism of LensCorruptions rests on two untested assumptions: setting the virtual pixel size by matching the sensor Nyquist frequency to the mean MTF20 value, and treating nominal lens prescriptions as representative of real lenses on the grounds that manufacturing tolerances are the only difference. If the MTF20 criterion is not appropriate, or if real lenses deviate substantially from their nominal designs, the severity matching and the 'realistic' qualification of LensCorruptions are weakened. Please add a sensitivity analysis, for example recomputing a subset of kernels under a different pixel-size criterion (MTF50, or a 2×MTF20 Nyquist choice) and comparing selected kernels against the toleranced PSF shown in the supplementary, to bound the effect on the reported accuracies and mAP values.
  3. [§VII-A, Tables VI and VIII] The OpticsAugment evaluation is partly circular: the augmentation is trained with kernels drawn from the same OpticsBench family on which it is evaluated, so the headline gains (about 18% average improvement on ImageNet-100 OpticsBench, and +7.7% mAP on MSCOCO-OpticsBench) are by construction to some degree. The transfer results to 2D common corruptions are encouraging and should be kept as independent evidence, but the paper does not evaluate OpticsAugment on LensCorruptions, which are the unseen realistic kernels. Please report the transfer to LensCorruptions, or an ablation that excludes the exact training kernels and severities, before concluding that the specific blur kernels can be compensated 'to a large extent.'
  4. [§VI-A and Suppl. F-A3, Fig. 25] The manuscript sometimes presents the disk baseline as uniquely weak, but the supplementary rank-correlation matrix shows that the OpticsBench corruptions are also only weakly correlated with each other (e.g. tau around 0.16–0.32 among astigmatism, coma, trefoil, defocus&spherical). If every single blur corruption is a weak proxy for others, the conclusion that a particular shaped kernel is needed should be phrased more carefully. Please either compare the disk baseline against another non-optical blur corruption (e.g. Gaussian blur) with the same analysis, or explicitly state that all single-kernel proxies are weak and the contribution is the availability of more diverse optical kernels.
minor comments (5)
  1. [Code footnote] The code URL contains a space ('classification robustness'); it should be URL-encoded or replaced with the repository's canonical URL.
  2. [Suppl. Table XXV vs Table IX] The signed improvements in Table XXV appear to use the opposite sign convention from Table IX: for many corruptions the cascade column is numerically higher than the OpticsAugment-only column while the delta column is negative, whereas Table IX reports positive 'additional average improvement.' Please define the delta explicitly and make the signs consistent between the main text and supplement.
  3. [§V-A] The sentence 'Besides depth, the simulation does not cover lens flare or scattering effects' is awkward; the first phrase should be clarified, for example by stating explicitly that depth-dependence is dropped for fixed-focus lenses and that other imaging effects are out of scope.
  4. [§VIII-A] The ImageNet-100 subset is attributed to reference [79] in a way that requires the reader to know the exact split from the Contrastive Multiview Coding paper; consider citing the specific ImageNet-100 split used or describing the selection procedure in one sentence.
  5. [Fig. 22] The simulator console output embedded in Fig. 22 is visually noisy and partly redundant with the plotted MTF curves; consider cropping the console text or moving it to the caption/appendix for readability.

Circularity Check

1 steps flagged · score 4.0 of 10

OpticsAugment's headline gains on OpticsBench are partly in-distribution by construction, but the central disk-vs-optical ranking claim is independently evaluated.

  1. fitted input called prediction [Sec. VII (OpticsAugment) and Sec. VII-A (Experiments on ImageNet-100); Tables VI, XXII-XXIII]
    "During dataloading in the training process each image is convolved with an individual RGB-kernel from the kernel stack containing e.g. 40 kernels for the different primary aberrations and severities. ... On top of the trained DNNs, all models are also trained with the same settings, but include OpticsAugment with a severity of 3 during training ... Table VI gives an overview of the improvement on ImageNet-100 OpticsBench with OpticsAugment."

    The OpticsAugment training distribution is exactly the OpticsBench test family: it samples the same 40 RGB kernels (same primary-aberration Zernike modes and severities) that define the OpticsBench corruptions, and the train/validation split is only over images, not over kernels or aberration types. The reported 18% average gain on ImageNet-100 OpticsBench (and the per-model gains in Tables VI, XXII-XXIII) therefore measures in-distribution recognition of the training augmentation, not a prediction about unseen optical degradations. The magnitude of the gain is forced by the overlap between the augmentation kernel stack and the evaluation kernel stack.

full rationale

The central claim that disk-shaped blur kernels are weak proxies for optical aberrations rests on rank correlations between the Hendrycks defocus baseline and OpticsBench/LensCorruptions; those correlations are empirical and not entailed by the construction, so that claim is not circular. The self-citations to [11] for OpticsBench and OpticsAugment are descriptive rather than load-bearing: the kernel equations, coefficient tables, and matching procedure are reproduced in the paper and supplement, so no unverified result is imported. The only construction-forced result is the OpticsAugment-on-OpticsBench gain, because training augmentation and test corruption share the same kernel stack. The separate concern that OpticsBench kernels differ from the disk baseline in both shape and chromatic structure is an experimental confound, not a derivational circularity, and is not scored as such. Overall score 4: one partial in-distribution claim in a secondary contribution, while the central benchmark comparison is independently evaluated.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The central claims rest on standard optics mathematics (Zernike expansion, Fourier PSF) and on domain assumptions about the fidelity of simulated nominal lens PSFs to real cameras. The free parameters are calibration choices: Zernike coefficients matched to an existing blur baseline, the OpticsAugment training severity, and the MSCOCO subset size. No new physical entities are introduced. The weakest point in the ledger is the assumption that nominal lens prescriptions plus the MTF20 pixel-size rule reproduce real optical behavior, which the paper itself supports only by argument, not by comparison to measured PSFs.

free parameters (3)
  • OpticsBench Zernike coefficients per severity = Matched to defocus blur baseline; values not listed in main text
    Section IV and supplementary E-B: for each of the four corruptions and five severities, Zernike coefficients are optimized by stepping in increments of 0.1 lambda so that MTF50, AUC, SSIM, and PSNR match the disk-shaped defocus blur baseline. These are fitted values, though fitted to a baseline kernel, not to model accuracy.
  • OpticsAugment training severity = 3
    Section VII-A: all ImageNet-100 OpticsAugment experiments train with severity 3. The choice is motivated as a reasonable medium severity, but it is a hand-chosen hyperparameter that directly affects the reported gains; no sweep or selection criterion is reported.
  • MSCOCO LensCorruptions subset size = 1000 images
    Section VIII-B: a 1000-image subset of the MSCOCO validation set is used to keep experiments feasible. The authors validate that the subset is similar to the full set (0.2% mAP difference), so the impact is small, but the subset size is a user-chosen parameter.
assumptions (7)
  • domain assumption Huygens PSF simulation correctly models wavefront propagation from lens prescriptions.
    Section III-C and supplementary C: the PSF is computed as the squared Fourier transform of the aberrated pupil function (Eq. 5). The accuracy of this model for photographic lenses is assumed from optics textbooks and the Zemax tool; no measured PSFs are used for validation.
  • domain assumption Space-variant lens blur can be approximated by a single convolution per small image region.
    Section III: "For small image regions, the space-variant lens blur can be approximated with a convolution". The paper applies one PSF per image, ignoring within-image PSF variation, which is a simplification for 224x224 crops.
  • domain assumption Depth dependence of the PSF is negligible beyond the hyperfocal distance.
    Section III: "Beyond the lens-dependent hyperfocal distance for fixed-focus lenses, the PSF varies only with angle and wavelength and the depth dependence can then be dropped. We assume this case here for simplicity."
  • domain assumption Nominal lens prescriptions from the public Reiley database are representative of real manufactured lenses.
    Section III-C: "We argue here that such measured lenses differ only in manufacturing tolerances, and therefore the nominal models from optical design give a good insight into real lenses."
  • domain assumption Setting the sensor Nyquist frequency to the mean MTF20 value of the lens is a valid pixel-size matching rule.
    Section III-B: "we set the Nyquist frequency of the system to the mean MTF20 value of different measurement points. Critically, this fixes an otherwise free parameter in our study, the pixel size." MTF20 is an industry figure of merit, but the specific rule is chosen for this study.
  • standard math Zernike polynomials form a complete and orthogonal basis for wavefront expansion.
    Supplementary C, Eq. 1-4: standard optics mathematics from Born and Wolf [1] and Zernike's original work [31]. This is accepted background.
  • domain assumption The convolution blur model ignores sensor noise, ISP, polarization, and color filter array effects.
    Section V-A Limitations: "we do not add a color filter array", "We do not use a sensor noise or image signal processor (ISP) model for simplicity". The authors state these are simplifications.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Examining the Impact of Optical Aberrations to Image Classification and Object Detection Models." pith.science (2026). https://pith.science/paper/MMNWDXDV

@misc{pith2026250418510,
  author       = {Pith},
  title        = {Pith review of: Examining the Impact of Optical Aberrations to Image Classification and Object Detection Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MMNWDXDV}},
  note         = {Machine review of arXiv:2504.18510}
}
read the original abstract

Deep neural networks (DNNs) have proven to be successful in various computer vision applications such that models even infer in safety-critical situations. Therefore, vision models have to behave in a robust way to disturbances such as noise or blur. While seminal benchmarks exist to evaluate model robustness to diverse corruptions, blur is often approximated in an overly simplistic way to model defocus, while ignoring the different blur kernel shapes that result from optical systems. To study model robustness against realistic optical blur effects, this paper proposes two datasets of blur corruptions, which we denote OpticsBench and LensCorruptions. OpticsBench examines primary aberrations such as coma, defocus, and astigmatism, i.e. aberrations that can be represented by varying a single parameter of Zernike polynomials. To go beyond the principled but synthetic setting of primary aberrations, LensCorruptions samples linear combinations in the vector space spanned by Zernike polynomials, corresponding to 100 real lenses. Evaluations for image classification and object detection on ImageNet and MSCOCO show that for a variety of different pre-trained models, the performance on OpticsBench and LensCorruptions varies significantly, indicating the need to consider realistic image corruptions to evaluate a model's robustness against blur.

Figures

Figures reproduced from arXiv: 2504.18510 by the authors.

Figure 1
Figure 1. Blur image corruptions applied to an ImageNet image. The effects [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. (Simplified view) Lens elements are lumped into a black box (left) and [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Lens selection: (a) 3D Zernike coefficient-space for the medium field [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (28 more)
Figure 5
Figure 5. Figure 5: Ranking for the best 50 models on ImageNet-1k with respect to the baseline [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Ranking for the 30 most robust models on ImageNet-1k using the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Average robustness on ImageNet-1k-OpticsBench in dependence of [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Cls (top) and Loc (bottom) errors from TIDE [69] for the different [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Accuracy on OpticsBench-ImageNet-100 for DNNs with (blue) [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Accuracy for ResNeXt50 evaluated on ImageNet-100-C 2D common corruptions w/wo OpticsAugment training and all severities 1-5 (circle, diamond, [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 11
Figure 11. Figure 11: Classification on ImageNet-100 lens corruptions for high and low [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Detection on the MSCOCO LensCorruptions for two high perform [PITH_FULL_IMAGE:figures/full_fig_p012_12.png]
Figure 13
Figure 13. Figure 13: To evaluate the difficulty of the MSCOCO subset used, we report [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]
Figure 14
Figure 14. Figure 14: Image processing scheme: The circular pupil (a) contains an aberrated [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Distribution of virtual pixel sizes for all lenses (blue) and the subset [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]
Figure 16
Figure 16. Figure 16: Distribution of general camera parameters: Effective Focal Length [PITH_FULL_IMAGE:figures/full_fig_p019_16.png]
Figure 18
Figure 18. Figure 18: These images can then be inferenced by a DNN for [PITH_FULL_IMAGE:figures/full_fig_p020_18.png]
Figure 17
Figure 17. Figure 17: 3D Zernike coefficient-space for different field positions illustrating the diversity of the lens aberrations. (a) 0.0, (b) 0.3, (c) 0.5, (d) 0.7 and (e) [PITH_FULL_IMAGE:figures/full_fig_p021_17.png]
Figure 20
Figure 20. Figure 20: Defocus blur from [7] for severities 1-5 used for kernel matching [PITH_FULL_IMAGE:figures/full_fig_p022_20.png]
Figure 18
Figure 18. Figure 18: Lens blur sampled at different distances from the lens center, [PITH_FULL_IMAGE:figures/full_fig_p022_18.png]
Figure 21
Figure 21. Figure 21: Kernels used to generate OpticsBench. Each row contains the [PITH_FULL_IMAGE:figures/full_fig_p022_21.png]
Figure 23
Figure 23. Figure 23: Comparison of coma corrruption with center of mass aligned to the [PITH_FULL_IMAGE:figures/full_fig_p023_23.png]
Figure 24
Figure 24. Figure 24: Four image examples from ImageNet-1k degraded by different OpticsBench blur types and severities. From left to right the severity increases from [PITH_FULL_IMAGE:figures/full_fig_p024_24.png]
Figure 25
Figure 25. Figure 25: Kendall tau rank correlation between all corruptions at severity [PITH_FULL_IMAGE:figures/full_fig_p025_25.png]
Figure 26
Figure 26. Figure 26: Ranking on OpticsBench-ImageNet-1k and baseline defocus [PITH_FULL_IMAGE:figures/full_fig_p025_26.png]
Figure 27
Figure 27. Figure 27: Ranking on OpticsBench-ImageNet-1k and baseline defocus [PITH_FULL_IMAGE:figures/full_fig_p026_27.png]
Figure 28
Figure 28. Figure 28: Ranking on OpticsBench-ImageNet-1k (reflective boundaries) and baseline defocus [PITH_FULL_IMAGE:figures/full_fig_p028_28.png]
Figure 29
Figure 29. Figure 29: OpticsBench: Flowers-102 (a) and Stanford Cars (b). Average [PITH_FULL_IMAGE:figures/full_fig_p030_29.png]
Figure 31
Figure 31. Figure 31: Cascading AugMix [26] and OpticsAugment for EfficientNet: [PITH_FULL_IMAGE:figures/full_fig_p031_31.png]
Figure 32
Figure 32. Figure 32: NuImages validation dataset mAP 50 for selected models. Figs. 33 and 34 show the ranking with the baseline (defo￾cus blur) for the selected object detection models on NuIm￾ages and all severities. Severities 1 and 2 are relatively similar to the baseline, but with a d…
Figure 33
Figure 33. Figure 33: Ranking comparisons on the NuImages validation dataset + Optics [PITH_FULL_IMAGE:figures/full_fig_p032_33.png]
Figure 34
Figure 34. Figure 34: Ranking comparisons on the NuImages validation dataset + Optics [PITH_FULL_IMAGE:figures/full_fig_p032_34.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

118 extracted references · 72 canonical work pages

  1. [1]

    Principles of optics: elec- tromagnetic theory of propagation, interference and diffraction of light

    Max Born and Emil Wolf. Principles of optics: elec- tromagnetic theory of propagation, interference and diffraction of light. 7th expanded ed. Cambridge ; New York: Cambridge University Press, 1999

    1999

  2. [2]

    Handbook of Optical Systems

    Herbert Gross, ed. Handbook of Optical Systems . 1st ed. V ol. 3. John Wiley & Sons, Ltd, 2006

    2006

  3. [3]

    Warren J. Smith. Modern optical engineering: the design of optical systems . 3rd ed. New York: McGraw Hill, 2000

    2000

  4. [4]

    Opto- Mechanical Systems Design, Two Volume Set

    Paul Yoder and Daniel Vukobratovich. Opto- Mechanical Systems Design, Two Volume Set . 0th ed. CRC Press, 2018

    2018

  5. [5]

    In collab

    Perspectives on Modern Optics and Imaging: With Practical Examples Zemax OpticStudio. In collab. with Ronian Siew. E-Book edition. 2023

    2023

  6. [6]

    Automotive mass production of camera systems: Linking image quality to AI perfor- mance

    Alexander Braun. “Automotive mass production of camera systems: Linking image quality to AI perfor- mance”. In: tm - Technisches Messen (2022)

    2022

  7. [7]

    Benchmark- ing Neural Network Robustness to Common Corrup- tions and Perturbations

    Dan Hendrycks and Thomas Dietterich. “Benchmark- ing Neural Network Robustness to Common Corrup- tions and Perturbations”. In: ICLR (2019)

    2019

  8. [8]

    Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming

    Claudio Michaelis et al. “Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming”. In: arXiv:1907.07484 [cs, stat] (2020)

    arXiv 2020

Show all 118 references
  1. [9]

    3D Common Corruptions and Data Augmentation

    O ˘guzhan Fatih Kar et al. “3D Common Corruptions and Data Augmentation”. en. In: CVPR. 2022

    2022

  2. [10]

    Examining the impact of blur on recognition by convolutional networks

    Igor Vasiljevic, Ayan Chakrabarti, and Gregory Shakhnarovich. “Examining the impact of blur on recognition by convolutional networks”. In: arXiv:1611.05760 (2016)

    2016 arXiv

  3. [11]

    Classification Robustness to Common Optical Aber- rations

    Patrick M ¨uller, Alexander Braun, and Margret Keuper. “Classification Robustness to Common Optical Aber- rations”. In: ICCV Workshops. 2023

    2023

  4. [12]

    URL: https://www.zemax

    OpticStudio — Optical, Illumination & Laser System Design Software - Zemax . URL: https://www.zemax. com/products/opticstudio (visited on 01/15/2023)

    2023

  5. [13]

    ImageNet Large Scale Visual Recognition Challenge

    Olga Russakovsky et al. “ImageNet Large Scale Visual Recognition Challenge”. In: IJCV 115.3 (2015)

    2015

  6. [14]

    Microsoft COCO: Common Ob- jects in Context

    Tsung-Yi Lin et al. “Microsoft COCO: Common Ob- jects in Context”. In: ECCV. 2014

    2014

  7. [15]

    nuScenes: A Multimodal Dataset for Autonomous Driving

    Holger Caesar et al. “nuScenes: A Multimodal Dataset for Autonomous Driving”. In: CVPR. 2020

    2020

  8. [16]

    Benchmarking Robustness of 3D Object Detection to Common Corruptions

    Yinpeng Dong et al. “Benchmarking Robustness of 3D Object Detection to Common Corruptions”. In: CVPR. 2023

    2023

  9. [17]

    Reliable eval- uation of adversarial robustness with an ensemble of diverse parameter-free attacks

    Francesco Croce and Matthias Hein. “Reliable eval- uation of adversarial robustness with an ensemble of diverse parameter-free attacks”. In: ICLR. 2020

    2020

  10. [18]

    DeepFool: A Simple and Ac- curate Method to Fool Deep Neural Networks

    Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. “DeepFool: A Simple and Ac- curate Method to Fool Deep Neural Networks”. In: CVPR. 2016

    2016

  11. [19]

    Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search

    Maksym Andriushchenko et al. “Square Attack: A Query-Efficient Black-Box Adversarial Attack via Random Search”. In: ECCV. 2020

    2020

  12. [20]

    Adversarial Ex- amples Are Not Easily Detected: Bypassing Ten De- tection Methods

    Nicholas Carlini and David Wagner. “Adversarial Ex- amples Are Not Easily Detected: Bypassing Ten De- tection Methods”. In: Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security. 2017

    2017

  13. [21]

    RobustBench: a standard- ized adversarial robustness benchmark

    Francesco Croce et al. “RobustBench: a standard- ized adversarial robustness benchmark”. In: Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track . 2021

    2021

  14. [22]

    Is RobustBench/AutoAttack a suit- able Benchmark for Adversarial Robustness?

    Peter Lorenz et al. “Is RobustBench/AutoAttack a suit- able Benchmark for Adversarial Robustness?” In: The AAAI-22 Workshop on Adversarial Machine Learning and Beyond. 2022

    2022

  15. [23]

    Improving robustness using gener- ated data

    Sven Gowal et al. “Improving robustness using gener- ated data”. In: NeurIPS 34 (2021)

    2021

  16. [24]

    ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness

    Robert Geirhos et al. “ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness”. In: ICLR. 2018

    2018

  17. [25]

    AutoAugment: Learning Aug- mentation Strategies From Data

    Ekin D. Cubuk et al. “AutoAugment: Learning Aug- mentation Strategies From Data”. In: CVPR. 2019

    2019

  18. [26]

    AugMix: A Simple Data Processing Method to Improve Robustness and Un- certainty

    Dan Hendrycks* et al. “AugMix: A Simple Data Processing Method to Improve Robustness and Un- certainty”. In: ICLR. 2020

    2020

  19. [27]

    The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generaliza- tion

    Dan Hendrycks et al. “The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generaliza- tion”. In: ICCV. 2021

    2021

  20. [28]

    Improving robustness against common corruptions with frequency biased models

    Tonmoy Saikia, Cordelia Schmid, and Thomas Brox. “Improving robustness against common corruptions with frequency biased models”. In: ICCV. 2021

    2021

  21. [29]

    Do adversarially robust imagenet models transfer better?

    Hadi Salman et al. “Do adversarially robust imagenet models transfer better?” In: NeurIPS 33 (2020)

    2020

  22. [30]

    ImageNet-Patch: A dataset for benchmarking machine learning robustness against ad- versarial patches

    Maura Pintor et al. “ImageNet-Patch: A dataset for benchmarking machine learning robustness against ad- versarial patches”. In: Pattern Recognition 134 (2023)

    2023

  23. [31]

    Beugungstheorie des schneidenver- fahrens und seiner verbesserten form, der phasenkon- trastmethode

    von F. Zernike. “Beugungstheorie des schneidenver- fahrens und seiner verbesserten form, der phasenkon- trastmethode”. In: Physica 1.7 (1934)

    1934

  24. [32]

    Retinal image quality for virtual eyes generated by a statistical model of ocular wave- front aberrations

    Larry N. Thibos. “Retinal image quality for virtual eyes generated by a statistical model of ocular wave- front aberrations”. In: Ophthalmic and Physiological Optics 29.3 (2009)

    2009

  25. [33]

    Optimal modeling of corneal surfaces with Zernike polynomi- 14 als

    D.R. Iskander, M.J. Collins, and B. Davis. “Optimal modeling of corneal surfaces with Zernike polynomi- 14 als”. In: IEEE Transactions on Biomedical Engineer- ing 48.1 (2001)

    2001

  26. [34]

    Lens design

    Milton Laikin. Lens design. 4th ed. Optical science and engineering 122. Boca Raton, FL: CRC Press, 2007

    2007

  27. [35]

    Direct determi- nation of aberration functions in microscopy by an artificial neural network

    Benjamin P. Cumming and Min Gu. “Direct determi- nation of aberration functions in microscopy by an artificial neural network”. In: Optics Express 28.10 (2020)

    2020

  28. [36]

    Calibration of quasi-static aberra- tions in exoplanet direct-imaging instruments with a Zernike phase-mask sensor

    M. N’Diaye et al. “Calibration of quasi-static aberra- tions in exoplanet direct-imaging instruments with a Zernike phase-mask sensor”. In: Astronomy & Astro- physics 555 (2013)

    2013

  29. [37]

    Wave-front reconstruction using a Shack–Hartmann sensor

    R. G. Lane and M. Tallon. “Wave-front reconstruction using a Shack–Hartmann sensor”. In: Applied Optics 31.32 (1992)

    1992

  30. [38]

    prysm: A Python optics module

    Brandon Dube. “prysm: A Python optics module”. In: Journal of Open Source Software 4.37 (2019)

    2019

  31. [39]

    3D PSF Models for Fluorescence Microscopy in ImageJ

    H. Kirshner, D. Sage, and M. Unser. “3D PSF Models for Fluorescence Microscopy in ImageJ”. In: Proceed- ings of the Twelfth International Conference on Meth- ods and Applications of Fluorescence Spectroscopy, Imaging and Probes (MAF’11) . 2011

    2011

  32. [40]

    Phillips and Henrik Eliasson

    Jonathan B. Phillips and Henrik Eliasson. Camera Im- age Quality Benchmarking. Newark, United Kingdom: John Wiley & Sons, Incorporated, 2018

    2018

  33. [41]

    Handbook of Optical Systems

    Herbert Gross, ed. Handbook of Optical Systems . 1st ed. V ol. 1. John Wiley & Sons, Ltd, 2005

    2005

  34. [42]

    Eugene Hecht. Optics. 5 ed/fifth edition, global edi- tion. Boston Columbus Indianapolis New York San Francisco Amsterdam Cape Town Dubai London Madrid Milan Munich: Pearson, 2017

    2017

  35. [43]

    Handbook of Optical Systems

    Herbert Gross, ed. Handbook of Optical Systems . 1st ed. V ol. 4. John Wiley & Sons, Ltd, 2008

    2008

  36. [44]

    Fast iterative image restoration with a spatially varying PSF

    James Nagy and Dianne O’Leary. “Fast iterative image restoration with a spatially varying PSF”. In: Optical Science, Engineering and Instrumentation. 1997

    1997

  37. [45]

    Joseph W. Goodman. Introduction to Fourier optics . Fourth edition. New York: W.H. Freeman, Macmillan Learning, 2017

    2017

  38. [46]

    Glenn D. Boreman. Modulation Transfer Function in Optical and Electro-Optical Systems . SPIE, 2001

    2001

  39. [47]

    Correcting Misleading Image Quality Measurements

    Norman Koren. “Correcting Misleading Image Quality Measurements”. In: Electronic Imaging (2020)

    2020

  40. [48]

    Digital Image Processing, Third Edition

    Rafael C. Gonzalez, Richard E. Woods, and Barry R. Masters. “Digital Image Processing, Third Edition”. In: Journal of Biomedical Optics 14.2 (2009)

    2009

  41. [49]

    Hand- buch Bauelemente der Optik

    Martin L ¨offler-Mang and Gottfried Schr ¨oder. Hand- buch Bauelemente der Optik. Ed. by Helmut Naumann. 8th ed. M ¨unchen: Hanser, 2020

    2020

  42. [50]

    URL: https : / / github

    vision/references/classification at v0.11.0 · pytorch/vision. URL: https : / / github . com / pytorch / vision/tree/v0.11.0/references/classification (visited on 03/01/2023)

    2023

  43. [51]

    ImageNet: A large-scale hierarchical image database

    Jia Deng et al. “ImageNet: A large-scale hierarchical image database”. In: CVPR. 2009

    2009

  44. [52]

    Lens-Designs

    Daniel Reiley. Lens-Designs. en-US. URL: https : / / www.lens-designs.com/ (visited on 08/10/2023)

    2023

  45. [53]

    3D-Printed Portable Robotic Mobile Microscope for Remote Diagnosis of Global Health Diseases

    Jaime Garc ´ıa-Villena et al. “3D-Printed Portable Robotic Mobile Microscope for Remote Diagnosis of Global Health Diseases”. In: Electronics 10.19 (2021)

    2021

  46. [54]

    URL: https://pytorch.org/vision/0.15/ models.html#classification (visited on 04/02/2024)

    Models and pre-trained weights — Torchvision 0.15 documentation. URL: https://pytorch.org/vision/0.15/ models.html#classification (visited on 04/02/2024)

    2024

  47. [55]

    CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features

    Sangdoo Yun et al. “CutMix: Regularization Strategy to Train Strong Classifiers With Localizable Features”. In: ICCV. 2019

    2019

  48. [56]

    mixup: Beyond Empirical Risk Minimization

    Hongyi Zhang et al. “mixup: Beyond Empirical Risk Minimization”. In: ICLR. 2018

    2018

  49. [57]

    Benjamin Erichson et al

    N. Benjamin Erichson et al. NoisyMix: Boosting Model Robustness to Common Corruptions . 2022. arXiv: 2202.01263[cs,stat]

    2022 arXiv

  50. [58]

    A ConvNet for the 2020s

    Zhuang Liu et al. “A ConvNet for the 2020s”. In: CVPR (2022)

    2022

  51. [59]

    Howard et al

    Andrew G. Howard et al. MobileNets: Efficient Con- volutional Neural Networks for Mobile Vision Appli- cations. 2017. arXiv: 1704.04861[cs]

    2017 arXiv

  52. [60]

    Deep Residual Learning for Image Recognition

    Kaiming He et al. “Deep Residual Learning for Image Recognition”. In: CVPR. 2016

    2016

  53. [61]

    Densely Connected Convolutional Networks

    Gao Huang et al. “Densely Connected Convolutional Networks”. In: CVPR. 2017

    2017

  54. [62]

    Aggregated Residual Transforma- tions for Deep Neural Networks

    Saining Xie et al. “Aggregated Residual Transforma- tions for Deep Neural Networks”. In: CVPR. 2017

    2017

  55. [63]

    Swin Transformer V2: Scaling Up Capacity and Resolution

    Ze Liu et al. “Swin Transformer V2: Scaling Up Capacity and Resolution”. In: CVPR. 2022

    2022

  56. [64]

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

    Mingxing Tan and Quoc Le. “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks”. In: ICLR. 2019

    2019

  57. [65]

    Vision Transformers for Dense Prediction

    Rene Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. “Vision Transformers for Dense Prediction”. In: ICCV. 2021

    2021

  58. [66]

    The Treatment of Ties in Ranking Problems

    M. G. Kendall. “The Treatment of Ties in Ranking Problems”. In: Biometrika 33.3 (1945)

    1945

  59. [67]

    Au- tomated Flower Classification over a Large Number of Classes

    Maria-Elena Nilsback and Andrew Zisserman. “Au- tomated Flower Classification over a Large Number of Classes”. In: 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing. 2008

    2008

  60. [68]

    3D Object Representations for Fine-Grained Categorization

    Jonathan Krause et al. “3D Object Representations for Fine-Grained Categorization”. In: 2013 IEEE Inter- national Conference on Computer Vision Workshops . 2013

    2013

  61. [69]

    TIDE: A General Toolbox for Identifying Object Detection Errors

    Daniel Bolya et al. “TIDE: A General Toolbox for Identifying Object Detection Errors”. In: ECCV. 2020

    2020

  62. [70]

    Cascade R- CNN: High Quality Object Detection and Instance Segmentation

    Zhaowei Cai and Nuno Vasconcelos. “Cascade R- CNN: High Quality Object Detection and Instance Segmentation”. In: IEEE Transactions on Pattern Analysis and Machine Intelligence 43.5 (2021)

    2021

  63. [71]

    Deformable DETR: Deformable Transformers for End-to-End Object Detection

    Xizhou Zhu et al. “Deformable DETR: Deformable Transformers for End-to-End Object Detection”. In: ICLR. 2020

    2020

  64. [72]

    DINO: DETR with Improved De- Noising Anchor Boxes for End-to-End Object Detec- tion

    Hao Zhang et al. “DINO: DETR with Improved De- Noising Anchor Boxes for End-to-End Object Detec- tion”. In: The Eleventh International Conference on Learning Representations. Sept. 29, 2022

    2022

  65. [73]

    Feature Pyramid Networks for Object Detection

    Tsung-Yi Lin et al. “Feature Pyramid Networks for Object Detection”. In: CVPR. 2017. 15

    2017

  66. [74]

    Mask R-CNN

    K. He et al. “Mask R-CNN”. In: IEEE ICCV. 2017

    2017

  67. [75]

    Focal Loss for Dense Object Detection

    Tsung-Yi Lin et al. “Focal Loss for Dense Object Detection”. In: ICCV. 2017

    2017

  68. [76]

    YOLOX: Exceeding YOLO Series in

    Zheng Ge et al. YOLOX: Exceeding YOLO Series in

  69. [77]

    Does Robustness on ImageNet Transfer to Downstream Tasks?

    Yutaro Yamada and Mayu Otani. “Does Robustness on ImageNet Transfer to Downstream Tasks?” In: CVPR. 2022

    2022

  70. [78]

    Proper Reuse of Image Classification Fea- tures Improves Object Detection

    Cristina Vasconcelos, Vighnesh Birodkar, and Vincent Dumoulin. “Proper Reuse of Image Classification Fea- tures Improves Object Detection”. In: CVPR. 2022

    2022

  71. [79]

    Contrastive Multiview Coding

    Yonglong Tian, Dilip Krishnan, and Phillip Isola. “Contrastive Multiview Coding”. In: ECCV. 2020

    2020

  72. [80]

    Emerging Properties in Self- Supervised Vision Transformers

    Mathilde Caron et al. “Emerging Properties in Self- Supervised Vision Transformers”. In: ICCV. 2021

    2021

  73. [81]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Alexey Dosovitskiy et al. “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale”. In: ICLR 2021. 2021

    2021

  74. [82]

    Smith et al

    Samuel L. Smith et al. ConvNets Match Vision Trans- formers at Scale . 2023. arXiv: 2310.16764[cs]

    2023 arXiv

  75. [83]

    Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Net- works

    Shaoqing Ren et al. “Faster R-CNN: Towards Real- Time Object Detection with Region Proposal Net- works”. In: NeurIPS. 2015

    2015

  76. [84]

    You Only Look Once: Unified, Real-Time Object Detection

    Joseph Redmon et al. “You Only Look Once: Unified, Real-Time Object Detection”. In: CVPR. 2016

    2016

  77. [85]

    YOLO-World: Real-Time Open- Vocabulary Object Detection

    Tianheng Cheng et al. YOLO-World: Real-Time Open- Vocabulary Object Detection . 2024. arXiv: 2401 . 17270[cs]

    2024

  78. [86]

    MMDetection: Open MMLab De- tection Toolbox and Benchmark

    Kai Chen et al. “MMDetection: Open MMLab De- tection Toolbox and Benchmark”. In: arXiv preprint arXiv:1906.07155 (2019)

    2019 arXiv

  79. [87]

    Grid R-CNN

    Xin Lu et al. “Grid R-CNN”. In: CVPR. 2019

    2019

  80. [88]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

    Ze Liu et al. “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows”. In: ICCV. 2021

    2021

  81. [89]

    YOLOv3: An Incre- mental Improvement

    Joseph Redmon and Ali Farhadi. YOLOv3: An Incre- mental Improvement. 2018. arXiv: 1804.02767[cs]

    2018 arXiv

  82. [90]

    Hybrid Task Cascade for Instance Segmentation

    Kai Chen et al. “Hybrid Task Cascade for Instance Segmentation”. In: CVPR. 2019

    2019

  83. [91]

    End-to-End Object Detection with Transformers

    Nicolas Carion et al. “End-to-End Object Detection with Transformers”. In: ECCV. 2020

    2020

  84. [92]

    FoveaBox: Beyound Anchor-Based Object Detection

    Tao Kong et al. “FoveaBox: Beyound Anchor-Based Object Detection”. In: IEEE Transactions on Image Processing 29 (2020)

    2020

  85. [93]

    FCOS: Fully Convolutional One-Stage Object Detection

    Zhi Tian et al. “FCOS: Fully Convolutional One-Stage Object Detection”. In: ICCV. 2019

    2019

  86. [94]

    Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adap- tive Training Sample Selection

    Shifeng Zhang et al. “Bridging the Gap Between Anchor-Based and Anchor-Free Detection via Adap- tive Training Sample Selection”. In: CVPR. 2020

    2020

  87. [95]

    Zernike polynomials: a guide

    Vasudevan Lakshminarayanan and Andre Fleck. “Zernike polynomials: a guide”. In: Journal of Modern Optics 58.7 (2011)

    2011

  88. [96]

    Simulating op- tical properties to access novel metrological parameter ranges and the impact of different model approxima- tions

    Patrick M ¨uller and Alexander Braun. “Simulating op- tical properties to access novel metrological parameter ranges and the impact of different model approxima- tions”. In: IEEE MetroAutomotive. 2022

    2022

  89. [97]

    Image Quality Assessment: From Error Visibility to Structural Similarity

    Z. Wang et al. “Image Quality Assessment: From Error Visibility to Structural Similarity”. In: IEEE Transactions on Image Processing 13.4 (2004)

    2004

  90. [98]

    IEEE Standard for Camera Phone Image Quality

    “IEEE Standard for Camera Phone Image Quality”. In: IEEE Std 1858-2016 (2017)

    2017

  91. [99]

    Standard

    ISO12233:2017, Photography — Electronic still pic- ture imaging — Resolution and spatial frequency re- sponses. Standard. Geneva, CH: ISO, 2017

    2017

  92. [100]

    Texture-based measurement of spatial frequency response using the dead leaves target: extensions, and application to real camera systems

    Jon McElvain et al. “Texture-based measurement of spatial frequency response using the dead leaves target: extensions, and application to real camera systems”. In: IS&T/SPIE Electronic Imaging. 2010

    2010

  93. [101]

    Refined measurement of digital image texture loss

    Peter D. Burns. “Refined measurement of digital image texture loss”. In: Electronic Imaging. 2013

    2013

  94. [102]

    URL: https : / / www.imatest.com/product/imatest- target- generator/ (visited on 03/01/2023)

    Imatest Target Generator — Imatest . URL: https : / / www.imatest.com/product/imatest- target- generator/ (visited on 03/01/2023)

    2023

  95. [103]

    BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning

    F. Yu et al. “BDD100K: A Diverse Driving Dataset for Heterogeneous Multitask Learning”. In: CVPR. 2020

    2020

  96. [104]

    As Large as It Gets – Studying Infinitely Large Convolutions via Neural Implicit Frequency Filters

    Julia Grabinski, Janis Keuper, and Margret Keuper. “As Large as It Gets – Studying Infinitely Large Convolutions via Neural Implicit Frequency Filters”. In: TMLR (Feb. 6, 2024)

    2024

  97. [105]

    vision/references/detection at main · pytorch/vision . GitHub. URL: https://github.com/pytorch/vision/tree/ main/references/detection (visited on 10/08/2023)

    2023

  98. [106]

    Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour

    Priya Goyal et al. Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour . 2018. arXiv: 1706 . 02677[cs]. Patrick M ¨uller is a postdoctoral researcher at University of Siegen with a focus on out-of-distribution robustness of computer vision algorithms to optical aberrati...

    2018

  99. [108]

    Overview: Table XIX lists the accuracies on the ImageNet-1k validation dataset without image corruption and the corresponding average accuracy on OpticsBench and Ta- (a) 0.0 0.2 0.4 0.6 0.0 0.2 0.4 0.6 0.8 1.0 1.2 mtf_curve_img_txt 0.0 0.1 0.2 0.3 0.4 0.5 0.0 0.2 0.4 0.6 0.8 1...

  100. [109]

    In Table II Swin v2 and ConvNeXt have the highest clean accuracies and also the highest accuracies on OpticsBench with a reduction of 31.3 % and 30.2 %

    Additional analysis: This paragraph provides a more detailed analysis of the model behaviour when confronted to optical aberrations. In Table II Swin v2 and ConvNeXt have the highest clean accuracies and also the highest accuracies on OpticsBench with a reduction of 31.3 % and...

  101. [110]

    When looking at different baselines, which is equivalent with flattening a specific corruption curve, other models will have ties to other corruptions

    Kendall tau rank correlation: At first glance, the Op- ticsBench corruptions seem to be closely tied to each other, however this perspective is not true in general. When looking at different baselines, which is equivalent with flattening a specific corruption curve, other mode...

  102. [111]

    The corruptions are sorted in both directions in the same way

    Since the ranking is symmetric, the upper triangle matrix is left blank intentionally. The corruptions are sorted in both directions in the same way. when all robust models are excluded from the computation. The rank correlation rises for most of the constellations. However, w...

  103. [112]

    26 and 27 show the ranking of the 72 image classification models with the baseline defocus blur [7] image corruption for all five severities on the ImageNet-1k OpticsBench

    Ranking results for more severities: Figs. 26 and 27 show the ranking of the 72 image classification models with the baseline defocus blur [7] image corruption for all five severities on the ImageNet-1k OpticsBench. In general, as the severity increases in Figs. 26 and 27, the...

  104. [113]

    XXI lists the six main errors as defined by Bolya et al

    TIDE errors: Tab. XXI lists the six main errors as defined by Bolya et al. in [69] on the validation dataset of MSCOCO and for OpticsBench (second column) the average over all corruptions and severities. While the classification error (Cls) doubles from 2.7 to 5.3, the localiz...

  105. [114]

    8 (top) for some models the Cls error depends on the image corruption, while for others it is quite similar

    Additional analysis: In Fig. 8 (top) for some models the Cls error depends on the image corruption, while for others it is quite similar. YOLO (c), Mask R-CNN (f) and in particular DINO (h) have similar Cls errors for all image corruptions. For the Faster R-CNN (a), Cascade R-...

  106. [115]

    Further analysis: The OpticsAugment trained ResNet50 on ImageNet-1k is reported to be robust to the primary aberrations. However, it only ranks 48/72 on the ImageNet- 1k validation set, which is largely due to the restricted standard training recipe of 90 epochs: the pre-train...

  107. [116]

    We first train the default baseline using the torchvision script as before with 90 epochs and no augmentations beyond random crop and flip

    Additional models: We here provide additional experi- ments on ImageNet-100 with different training strategies for ResNet50 models and report the results in Table XXIV. We first train the default baseline using the torchvision script as before with 90 epochs and no augmentatio...

  108. [117]

    31 shows a comparison of EfficientNet trained with OpticsAugment (red) and a cascaded application of OpticsAugment & AugMix (blue) evaluated on 2D common corruptions [7]

    Cascading AugMix & OpticsAugment: Fig. 31 shows a comparison of EfficientNet trained with OpticsAugment (red) and a cascaded application of OpticsAugment & AugMix (blue) evaluated on 2D common corruptions [7]. The results are also listed in Table XXV. In general, the cascading...

  109. [118]

    First, we train a Faster R-CNN with an ImageNet pre-trained ResNet50 backbone applied as a Feature Pyramid Network (FPN) [73] as a baseline

    Training with OpticsAugment on MSCOCO: Ta- ble XXVI lists different variants for training a Faster R-CNN on MSCOCO with OpticsAugment. First, we train a Faster R-CNN with an ImageNet pre-trained ResNet50 backbone applied as a Feature Pyramid Network (FPN) [73] as a baseline. W...

  110. [2021]

    arXiv: 2107.08430[cs]

    2021. arXiv: 2107.08430[cs]

    2021 arXiv

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.