Pith. sign in

REVIEW 4 major objections 6 minor 49 references

Err on the Side of Texture: Texture Bias on Real Data

T0 review · 4 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Texture bias—models' reliance on repeated surface patterns rather than object shape—explains the confident mispredictions in ImageNet-A, where over 90% of these natural adversarial examples contain textures misaligned with their true label.

desk verdict Useful TAV/TID methods paper whose central ImageNet-A explanation is a self-consistency loop; publish the method, not the causal claim. read the letter →

arxiv 2412.10597 v2 pith:32CVXENT submitted 2024-12-13 cs.CV cs.CR

Blaine Hoak , Ryan Sheatsley , Patrick McDaniel This is my paper · ORCID
classification cs.CVcs.CR
keywords texturebiasnaturaladversarialexamplesAssociationValueidentificationImageNet-Amodelrobustnessconfidencecomputervision
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 sets out to show that texture bias—the tendency of image classifiers to rely on repeated surface patterns rather than object shape—is a dominant force in real-world behavior, and that it explains why natural adversarial examples exist. Natural adversarial examples are ordinary photographs that models confidently misclassify. The authors introduce the Texture Association Value (TAV), a score for how strongly a model links each of 56 texture classes to each of 1,000 object classes, and the Texture Identification (TID) method, which tags a real image with the texture whose expected model response it most resembles. They find that models confidently classify isolated textures, that accuracy and confidence on ImageNet validation images are dramatically higher when the image contains the dominant texture of the true label or predicted class, and that over 90% of ImageNet-A samples contain a texture that disagrees with the dominant texture of their true label. If correct, the work means texture misalignment is a sufficient cause of confident wrong predictions on real data, redirecting robustness research toward texture rather than only toward adversarial perturbations.

What carries the argument

The central object is the Texture Association Value (TAV), an $n \times m$ matrix where $n=56$ texture classes and $m=1000$ ImageNet object classes, computed from model argmax predictions on the Prompted Textures Dataset. Each entry is the product of four factors: the probability that a texture is predicted as an object class, the complement of the entropy of that texture's prediction distribution, the probability that predictions of an object class come from that texture, and the complement of the entropy of that object class's texture distribution. The companion Texture Identification (TID) assigns a texture to a real image by cosine similarity between the image's softmax output and the TAV rows. This machinery lets the authors detect which texture a model 'sees' in a real image without additional training, and then relate that texture to the model's accuracy and confidence.

What would settle it

A concrete test: take ImageNet-A images that the paper labels as misaligned (texture agrees with the prediction, disagrees with the label) and apply a style-transfer or texture-replacement method that changes the image's texture to the dominant texture of its true label while preserving the object's shape. If texture bias is the cause, these images should flip to correct predictions with high confidence. If a large fraction remain confidently misclassified after the texture is aligned, the texture-misalignment explanation is insufficient.

Watch

Extended reading notes

Core claim

The central discovery is that the confident, incorrect predictions collected in ImageNet-A are explained by texture bias. The paper argues that models learn strong associations between specific textures and object classes, that these associations drive predictions on real images, and that natural adversarial examples are precisely the cases where the image's dominant texture agrees with the predicted (wrong) class but disagrees with the true label. The authors demonstrate this by showing that models make confident predictions on isolated texture images, that accuracy and confidence on the ImageNet validation set are substantially higher when the dominant texture of the true label (or predicted class) is present, and that more than 90% of ImageNet-A samples have textures misaligned with their true label's dominant texture. They conclude that texture bias is a sufficient condition for the existence of natural adversarial examples.

Load-bearing premise

The load-bearing premise is that the Texture Identification (TID) step correctly labels the texture present in a real image, because the entire ImageNet-A analysis compares the texture the model responds to against the texture dominant for each class; if TID is wrong about what texture is present, the claim that over 90% of these images contain a misaligned texture collapses. The paper's own human validation reaches only 61% agreement, with some texture classes near chance levels, so this premise is weakly supported.

Editorial extensions

If this is right

  • If a model's accuracy on an object class depends on whether that class's dominant texture is present, then evaluation sets that happen to lack that texture will systematically understate the model's real-world accuracy.
  • Natural adversarial examples could be curated or generated by selecting images whose dominant texture matches a wrong class and mismatches the true class, giving a principled way to build robustness test sets.
  • Mitigation strategies for texture bias should target the learned texture-object associations themselves, rather than only adding shape cues during training.
  • The TAV metric can be used as a pre-deployment audit tool: list the textures each class relies on and flag classes where the dominant texture is an unexpected shortcut such as a grid for window screens.
  • Confident mispredictions on ordinary photographs may not require an adversary; many may be ordinary texture-driven errors.

Reading between the lines

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

  • Beyond the paper: the same TAV/TID machinery could be applied to other visual dimensions such as color, lighting, or background scene by substituting an appropriate dataset, as the authors themselves note for color.
  • The 61% human agreement rate on TID suggests texture identification is inherently ambiguous; a stricter test of the causal claim would manipulate the texture in an image while keeping shape fixed and show predictions flip accordingly.
  • If texture misalignment is the mechanism, then adversarial perturbations that inject the dominant texture of a wrong class should be especially effective, connecting texture bias to classical adversarial attacks in a testable way.
  • The paper's single-dominant-texture assumption may cause TID to mislabel images with mixed textures; a multi-label extension of TID could strengthen or qualify the ImageNet-A result.
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 the Texture Association Value (TAV), a metric computed from a model's softmax predictions on the Prompted Textures Dataset, and the Texture Identification (TID) procedure, which assigns a texture class to a real image by cosine similarity between the model's softmax output and TAV rows. The authors use TID to analyze texture bias on ImageNet validation images and to study ImageNet-A natural adversarial examples. They report that model accuracy and confidence on ImageNet validation depend strongly on whether the TID-identified texture is the dominant texture for the object class, and they claim that over 90% of ImageNet-A samples contain textures misaligned with their true label but aligned with the prediction, concluding that texture bias explains confident mispredictions. Experiments are run on eight torchvision models, with the main results reported for ResNet50.

Significance. If the central interpretive claims were supported, the paper would be a valuable extension of texture-bias research from synthetic cue-conflict data to real-world images, and it would offer a potential explanation for a robustness phenomenon. The paper has concrete strengths: the TAV/TID framework is clearly specified, the code is released, the evaluation spans several architectures, and the authors undertake a human validation study of TID. However, the headline result is that texture misalignment explains the existence of natural adversarial examples, and that claim is not supported by the current experimental design. The analysis depends on model-derived texture labels, and the human validation is too weak and too indirect to break this dependency. The paper is worth revising, but the causal and explanatory framing must be substantially reworked or supplemented with independent evidence.

major comments (4)
  1. [Section III-B, Eq. (3); Section IV-D, Figures 12–13] The central claim that over 90% of ImageNet-A samples contain textures misaligned with the true label is largely a formal consequence of how TID, prediction textures, and label textures are constructed. TID labels an image by maximizing cosine similarity between the model's softmax output and TAV rows, and TAV rows are themselves built from the same model's softmax predictions (Eq. 2). The prediction texture for a class is the TID mode over validation images predicted as that class (Figure 10). An ImageNet-A image that is confidently predicted as class j therefore has a softmax vector similar to validation images predicted as j, so its TID label will tend to match the prediction texture and not the label texture. The human evaluation in Section III-B1 does not break this loop, because it was run on ImageNet validation images and did not independently verify the label/prediction alignment statistic. The explanatory conclusion 'Natural adversarial examples are a consequence of texture bias' (Section IV-D takeaway) and the Abstract's phrasing 'texture bias explains the existence' are therefore not supported by the statistics as reported. Independent texture annotations for ImageNet-A or a matched control group of correctly classified images is needed.
  2. [Section III-B1, Figure 4] The human validation of TID reaches only 61% overall agreement, and the agreement varies widely by texture class, with classes near chance (the text reports 'cobwebbed' at 13%, and Figure 4 shows several classes below 40%). This level of accuracy is insufficient to establish that TID reliably identifies textures in the ImageNet-A analysis, especially because the alignment statistic depends on correct texture labels for individual images. The paper should report per-class agreement for all 56 texture classes with sample counts, and it should test whether the ImageNet-A alignment result in Figure 12 is robust when only high-agreement texture classes are used.
  3. [Section IV-D, Figures 12–13] The causal claim that texture misalignment explains confident mispredictions requires a control group, but the paper analyzes only ImageNet-A misclassified images. Without showing that correctly classified images overwhelmingly contain their label's dominant texture, and that the label-texture mismatch rate is significantly lower for correct predictions than for ImageNet-A mispredictions, the reported rates cannot distinguish a texture-misalignment cause from a generic property of the TID assignment. The paper also states that texture misalignment is a 'sufficient condition' (Section III-B and IV-D) for confident misprediction, but it never provides evidence of sufficiency, such as an intervention that changes a predicted class by changing texture while holding shape constant.
  4. [Section III-B, Eq. (3)] The TID forces each image to a single texture class via argmax over 56 classes, but the paper acknowledges in Section III-B1 that natural images often contain multiple overlapping textures. This assumption could systematically bias the alignment statistic: an image whose model response is driven by a background texture will be assigned that texture even if the foreground object exhibits a different one. The authors should quantify the sensitivity of the ImageNet-A alignment results to this assumption, for example by thresholding the cosine similarity or reporting the full similarity distribution rather than only the argmax.
minor comments (6)
  1. [Abstract] The abstract contains a formatting artifact ('Texture Association V alue') that should be corrected.
  2. [Section IV-C] There is a duplicated word in the text: 'shows shows' appears in the discussion of Figure 8.
  3. [Section IV-D] The phrase 'the magitude of the similarity' contains a typo; it should be 'magnitude'.
  4. [Section IV-D, Figure 12] The text reports 'over 90% of the samples contain textures that disagree with the texture associated with their true label' and points to the blue and orange bars, but the decomposition of the four categories into the >90% figure is not stated explicitly. The authors should define the exact union of categories used in the claim.
  5. [Section IV-C, Figures 9 and 11] Accuracy and confidence differences between dominant and non-dominant textures are reported as point estimates without error bars or statistical significance tests; given the grouped structure of the data (images nested in labels), a mixed-effects or per-label paired analysis would strengthen these claims.
  6. [Section IV-D, Figure 13] Figure 13 reports per-class agreement rates but many classes contain very few ImageNet-A samples (the paper mentions a class with a single sample). Displaying sample counts or confidence intervals per class would prevent overinterpretation of individual bars.

Circularity Check

2 steps flagged · score 7.0 of 10

ImageNet-A 'texture misalignment' claim is a self-consistency loop: TID, dominant class textures, and the claimed 'texture in the image' are all computed from the same model softmax, so the >90% label mismatch is forced by ImageNet-A's confident-misprediction selection criterion rather than by measured texture content.

  1. self definitional [Section III-B, Eq. 3, and Section IV-D ('prediction texture' definition via Figure 10)]
    "The TID works by comparing the model’s softmax outputs for a real image to each row in the TAV matrix (representing each texture) and selecting the texture with the highest similarity to the image’s output distribution. More formally, the TID for an image is calculated as: TID(x) = arg max_i softmax(fθ(x)) · TAV_i / ∥softmax(fθ(x))∥ · ∥TAV_i∥ (3) ... using the object class that each natural adversarial example is predicted as, we get the prediction texture by identifying the most dominant texture from the ImageNet data (from the upper envelope of Figure 10) for that object class."

    The 'texture present in the image' is not measured independently; it is defined as the TAV row (itself built from the same model's predictions on PTD, Eq. 2) that maximizes cosine similarity with the image's own softmax vector. The 'dominant texture for the prediction class' is the mode of these same TID labels over validation images predicted as that class. For a confidently mispredicted ImageNet-A sample, whose softmax is peaked at the predicted class j, the TID label is selected from the very same rows that define class-j's texture profile. The reported agreement between identified texture and prediction texture is therefore a formal consequence of shared definitions, not independent evidence about the image's visual texture.

  2. fitted input called prediction [Section IV-D, Figure 12 analysis and abstract claim]
    "In Figure 12 we show the ratio of total samples in the ImageNet-A dataset that contain a texture that agree with their prediction texture, label texture, neither, and both. ... More than 90% of the samples contain textures that disagree with the texture associated with their true label (i.e., samples in the blue and orange bars). We find that the presence of this different, misaligned texture explains natural adversarial examples’ confident mispredictions."

    ImageNet-A is defined as a set of inputs that the model confidently misclassifies. Each sample's softmax is therefore, by the dataset's selection criterion, not aligned with the model's characteristic response to the true label. The TID labels the image using that same softmax, and the 'label texture' is the TID mode over validation images of the true label. Thus the >90% label mismatch is a restatement of the selection criterion used to build ImageNet-A, relabeled as a texture-content measurement. The statistic would be largely forced even if the images visually contained the true label's dominant texture, because the model's misdirected output, not the image's actual texture, determines the TID label.

full rationale

The paper's TAV and TID are coherent descriptive statistics of model behavior: TAV summarizes how often PTD textures are classified as each ImageNet class, and the confidence histograms on PTD are independent empirical observations. The circularity enters at the ImageNet-A step. The quantity called 'the texture present in the image' is defined by Eq. 3 as the TAV row whose model-response profile is closest to the image's own softmax; the 'dominant texture for an object class' is the TID mode over validation images predicted as that class. Both variables are functions of the same model's output distribution. Since ImageNet-A is selected as confidently misclassified data, each sample's softmax is peaked at a non-label class, so the TID label tends to match the prediction-class texture profile and fail to match the label-class texture profile by construction. The reported >90% label mismatch and high prediction-texture agreement in Figure 12 are a restatement of the dataset's selection criterion rather than evidence that the images' visual textures differ from the true label's learned texture. The human evaluation (61% agreement) supports the TID as a rough texture-labeling heuristic on ImageNet validation images, but it was not run on ImageNet-A and does not validate the specific label/prediction alignment statistic. The central explanatory claim is therefore partially circular; the earlier descriptive analyses are not. No load-bearing self-citation chain was found; the PTD citation is a dataset source rather than an imported uniqueness theorem.

Assumptions & free parameters 0 free parameters · 6 assumptions · 2 invented entities

The paper introduces no numeric free parameters, but it relies on several domain assumptions about the validity of synthetic textures as probes of learned texture associations, the single-texture-per-image simplification, and the generalizability of per-class dominance statistics. The TAV formula itself is an ad hoc construction. The main invented constructs are the TAV and TID, both of which are evaluated only through the paper's own pipeline or a modest human study.

assumptions (6)
  • domain assumption Model output distributions on synthetic texture images faithfully represent the model's learned texture associations.
    TAV is constructed from model predictions on PTD texture images (Section III-A). If these synthetic textures do not match the textures the model learned from ImageNet, the associations are unrepresentative.
  • domain assumption The texture present in a real image can be identified by cosine similarity between the model's softmax output and TAV rows.
    TID relies on this assumption (Eq. 3). The human evaluation supports it only partially, with 61% overall agreement and chance-level agreement for some texture classes.
  • domain assumption Each image contains a single dominant texture that can be labeled.
    TID assigns exactly one texture per image. The paper acknowledges in the human evaluation discussion that real images often contain multiple overlapping textures.
  • domain assumption The Prompted Textures Dataset covers the space of textures that ImageNet-trained models have learned.
    The paper states in Limitations that the method relies on available texture data and may miss textures whose structure is not present in existing texture datasets.
  • domain assumption ImageNet validation per-class dominant textures generalize to ImageNet-A.
    Section IV-D derives expected texture classes from ImageNet validation data and applies them to ImageNet-A images, which are out-of-distribution.
  • ad hoc to paper The entropy-complement weighting in the TAV formula is a valid way to combine association strength.
    Equation 2 multiplies probabilities by entropy complements without a derivation or external justification; it is a hand-designed formula.
invented entities (2)
  • Texture Association Value (TAV)
    purpose: A score matrix quantifying the association between 56 texture classes and 1000 object classes based on model predictions.
    TAV is defined in Eq. 2 and is only indirectly validated through the TID human study. There is no independent benchmark that establishes TAV as a ground-truth measure of model texture reliance.
  • Texture Identification (TID) independent evidence
    purpose: A procedure that assigns a texture class to a real image by cosine similarity between the model's softmax output and TAV rows.
    TID produces falsifiable texture labels that can be compared against human judgments or independent texture segmentation. The paper includes a human evaluation on ImageNet validation, but with modest agreement (61% overall).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Err on the Side of Texture: Texture Bias on Real Data." pith.science (2026). https://pith.science/paper/32CVXENT

@misc{pith2026241210597,
  author       = {Pith},
  title        = {Pith review of: Err on the Side of Texture: Texture Bias on Real Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/32CVXENT}},
  note         = {Machine review of arXiv:2412.10597}
}
read the original abstract

Bias significantly undermines both the accuracy and trustworthiness of machine learning models. To date, one of the strongest biases observed in image classification models is texture bias-where models overly rely on texture information rather than shape information. Yet, existing approaches for measuring and mitigating texture bias have not been able to capture how textures impact model robustness in real-world settings. In this work, we introduce the Texture Association Value (TAV), a novel metric that quantifies how strongly models rely on the presence of specific textures when classifying objects. Leveraging TAV, we demonstrate that model accuracy and robustness are heavily influenced by texture. Our results show that texture bias explains the existence of natural adversarial examples, where over 90% of these samples contain textures that are misaligned with the learned texture of their true label, resulting in confident mispredictions.

Figures

Figures reproduced from arXiv: 2412.10597 by the authors.

Figure 1
Figure 1. ImageNet-A [6] examples misclassified as honeycombs on ResNet50. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Images from the ImageNet validation set identified as having grid [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Average agreement with human evaluators and number of samples evaluated for each predicted texture class. Horizontal line shows the overall [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figures from the paper (28 more)
Figure 5
Figure 5. Figure 5: Samples labeled as having a “swirly” texture by human evaluators [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Confidence histogram of the classification of texture images on [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Top 50 object-texture class pairs with the highest [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Scatter plot of the texture groupings present in each label by how many samples are in each group (normalized by number of samples in each label). [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: The accuracy of samples that do and do not contain the dominant [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: Scatter plot of the texture groupings present in each object prediction by how many samples are in each group (normalized by total number of [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: The average confidence of samples that do and do not contain the [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: The average alignment between the identified and the most common [PITH_FULL_IMAGE:figures/full_fig_p010_12.png]
Figure 13
Figure 13. Figure 13: The rate of agreement between the textures identified in ImageNet-A images and the textures predominantly found in the respective object labels and [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]
Figure 14
Figure 14. Figure 14: Mean TID magnitude for ImageNet validation data and ImageNet-A. [PITH_FULL_IMAGE:figures/full_fig_p011_14.png]
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 15
Figure 15. Figure 15: Images identified by TID as having a checkered texture. [PITH_FULL_IMAGE:figures/full_fig_p016_15.png]
Figure 16
Figure 16. Figure 16: Images identified by TID as having a scaly texture. [PITH_FULL_IMAGE:figures/full_fig_p016_16.png]
Figure 19
Figure 19. Figure 19: Images identified by TID as having a perforated texture. [PITH_FULL_IMAGE:figures/full_fig_p017_19.png]
Figure 21
Figure 21. Figure 21: Confidence histogram of texture images on ResNet152. [PITH_FULL_IMAGE:figures/full_fig_p017_21.png]
Figure 22
Figure 22. Figure 22: Confidence histogram of texture images on ConvNeXT. [PITH_FULL_IMAGE:figures/full_fig_p017_22.png]
Figure 23
Figure 23. Figure 23: Confidence histogram of texture images on Inception-v3. [PITH_FULL_IMAGE:figures/full_fig_p018_23.png]
Figure 27
Figure 27. Figure 27: Top 50 strongest TAV pairs on ResNet18. coil - spiraled waffle iron - waffled knot - braided wig - wavy cardigan - knitted vestment - paisley paintbrush - marbled window screen - grid gown - pleated coil - swirly paintbrush - swirly mask - freckled chime - striped sea…
Figure 29
Figure 29. Figure 29: Top 50 strongest TAV pairs on ConvNeXT [PITH_FULL_IMAGE:figures/full_fig_p018_29.png]
Figure 35
Figure 35. Figure 35: ResNet152. water jug space bar packet monitor black-and-tan coonhound tub bolete cradle Appenzeller swing bicycle-built-for-two Great Dane tape player hand-held computer suspension bridge tabby pop bottle basset garbage truck dining table red wolf bookcase American ch…
Figure 36
Figure 36. Figure 36: ConvNeXT. pickup thresher mountain tent CD player motor scooter Greater Swiss Mountain dog hot pot screwdriver Border terrier miniskirt tailed frog tape player shopping basket Bernese mountain dog combination lock pop bottle hognose snake Irish terrier frying pan gard…
Figure 33
Figure 33. Figure 33: Top 50 strongest TAV pairs on DenseNet169. ladle toilet tissue confectionery notebook caldron water jug cleaver jeep redbone fountain pen monastery mortarboard cairn overskirt beaker bloodhound giant schnauzer kelpie home theater Japanese spaniel Bernese mountain dog …
Figure 34
Figure 34. Figure 34: ResNet18. Newfoundland jeep tow truck coyote horse cart cash machine Dungeness crab Chihuahua wool pickelhaube Sealyham terrier horizontal bar chocolate sauce weasel plastic bag pier loupe Madagascar cat Lhasa microphone affenpinscher barbershop miniature schnauzer su…
Figure 41
Figure 41. Figure 41: ResNet18. moving van reflex camera garbage truck Chesapeake Bay retriever bull mastiff red wolf trombone combination lock pickelhaubepug knee pad cradle English springer toyshop electric locomotive hot pot screwdriver jinrikisha cannon night snake microphone oxcart Si…
Figure 42
Figure 42. Figure 42: ResNet152. Newfoundland sliding door siamang bookshop Greater Swiss Mountain dog reflex camera rock python clumber langur boxer prison wire-haired fox terrier ice cream red wine lumbermill electric locomotive bottlecap monastery cleaver carbonara common iguana bucket …
Figure 43
Figure 43. Figure 43: ConvNeXT. pickup mountain tent minibus car wheel hot pot modem Border terrier chain screen abaya space bar ice cream marmot garden spider pill bottlepot analog clock parallel bars tailed frog joystick hognose snake scalepier espresso carton ski bonnet oil filter pedes…
Figure 45
Figure 45. Figure 45: EfficientNet-B0. whiskey jug ashcan fox squirrel moped shoe shop Cardigan tailed frog dining table Appenzeller wing bikini hot pot dowitcher German shepherd water bottle minibus anemone fish folding chair grasshopper loudspeaker four-poster soft-coated wheaten terrier…
Figure 46
Figure 46. Figure 46: DenseNet121. sliding door moped screwdriver ground beetle cradle CD player sunglasses Australian terrier hammer long-horned beetle notebook English foxhound restaurant wing pole agaric plastic bag hot pot flat-coated retriever spider monkey velvet wooden spoon stethos…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 19 canonical work pages

  1. [1]

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

    R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wichmann, and W. Brendel, “ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness,” en, in International Conference on Learning Representations (ICLR) , Jan

    work page

  2. [2]

    On the Performance of GoogLeNet and AlexNet Applied to Sketches,

    P. Ballester and R. Araujo, “On the Performance of GoogLeNet and AlexNet Applied to Sketches,” en, Proceedings of the AAAI Conference on Artificial In- telligence, vol. 30, no. 1, Feb. 2016, Number: 1, ISSN : 2374-3468. DOI: 10.1609/aaai.v30i1.10171. [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/ view/10171

    doi:10.1609/aaai.v30i1.10171 2016

  3. [3]

    Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet,

    W. Brendel and M. Bethge, “Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet,” in International Conference on Learning Representations (ICLR) 2019 , arXiv, Mar. 2019. DOI: 10.48550/arXiv.1904.00760. [Online]. Available: http: //arxiv.org/abs/1904.00760

    arXiv doi:10.48550/arxiv.1904.00760 2019

  4. [4]

    P. Chen, C. Agarwal, and A. Nguyen, The shape and simplicity biases of adversarially robust ImageNet- trained CNNs, Sep. 2022. DOI: 10.48550/arXiv.2006. 09373. [Online]. Available: http://arxiv.org/abs/2006. 09373

    doi:10.48550/arxiv.2006 2022

  5. [5]

    Generalisation in humans and deep neural networks,

    R. Geirhos, C. R. M. Temme, J. Rauber, H. H. Schütt, M. Bethge, and F. A. Wichmann, “Generalisation in humans and deep neural networks,” in Conference on Neural Information Processing Systems (NeurIPS) 2018 , arXiv, Oct. 2020. DOI: 10.48550/arXiv.1808.08750. [Online]. Available: http://arxiv.org/abs/1808.08750

    arXiv doi:10.48550/arxiv.1808.08750 2018

  6. [6]

    Natural Adversarial Examples,

    D. Hendrycks, K. Zhao, S. Basart, J. Steinhardt, and D. Song, “Natural Adversarial Examples,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2021, arXiv, Mar. 2021. DOI: 10.48550/arXiv. 1907.07174. [Online]. Available: http://arxiv.org/abs/ 1907.07174

    arXiv doi:10.48550/arxiv.1907.07174 2021

  7. [7]

    On Synthetic Texture Datasets: Challenges, Creation, and Curation

    B. Hoak and P. McDaniel, On Synthetic Texture Datasets: Challenges, Creation, and Curation , Sep. 2024. DOI: 10.48550/arXiv.2409.10297. [Online]. Available: http: //arxiv.org/abs/2409.10297

    work page Pith review arXiv doi:10.48550/arxiv.2409.10297 2024

  8. [8]

    Explorations in Texture Learning

    B. Hoak and P. McDaniel, “Explorations in Texture Learning,” in International Conference on Learning Representations (ICLR) 2024, Tiny Papers Track , arXiv, Mar. 2024. DOI: 10.48550/arXiv.2403.09543. [Online]. Available: http://arxiv.org/abs/2403.09543

    work page Pith review arXiv doi:10.48550/arxiv.2403.09543 2024

Show all 49 references
  1. [9]

    Describing Textures in the Wild,

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi, “Describing Textures in the Wild,” en, in 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, OH, USA: IEEE, Jun. 2014, pp. 3606–3613, ISBN : 978-1-4799-5118-5. DOI: 10.1109/ CVPR.2014.461. [On...

    2014

  2. [10]

    Shortcut Learning in Deep Neural Networks,

    R. Geirhos, J.-H. Jacobsen, C. Michaelis, et al., “Shortcut Learning in Deep Neural Networks,” Nature Machine Intelligence, vol. 2, no. 11, pp. 665–673, Nov. 2020, ISSN : 2522-5839. DOI: 10.1038/s42256-020-00257-z. [Online]. Available: http://arxiv.org/abs/2004.07780

    2020 arXiv

  3. [11]

    Intriguing properties of neural networks,

    C. Szegedy, W. Zaremba, I. Sutskever, et al., “Intriguing properties of neural networks,” in International Confer- ence on Learning Representations (ICLR) 2014 , arXiv, Feb. 2014. DOI: 10.48550/arXiv.1312.6199. [Online]. Available: http://arxiv.org/abs/1312.6199

    2014 doi arXiv

  4. [12]

    Evasion Attacks against Machine Learning at Test Time,

    B. Biggio, I. Corona, D. Maiorca, et al. , “Evasion Attacks against Machine Learning at Test Time,” in ECML PKDD, 2013. DOI: 10.1007/978-3-642-40994-3_

    2013 doi

  5. [13]

    Explaining and Harnessing Adversarial Examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and Harnessing Adversarial Examples,” in International Conference on Learning Representations (ICLR) 2015 , arXiv, Mar. 2015. DOI: 10 . 48550 / arXiv. 1412 . 6572. [Online]. Available: http://arxiv.org/abs/1412.6572

    2015 doi arXiv

  6. [14]

    Towards Deep Learning Models Resistant to Adversarial Attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards Deep Learning Models Resistant to Adversarial Attacks,” in International Conference on Learning Representations (ICLR) 2018, arXiv, Sep. 2019. DOI: 10.48550/arXiv.1706.06083. [Online]. Available: http://arxiv...

    2018 doi arXiv

  7. [15]

    DeepFool: A simple and accurate method to fool deep neural networks,

    S.-M. Moosavi-Dezfooli, A. Fawzi, and P. Frossard, “DeepFool: A simple and accurate method to fool deep neural networks,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2016 , arXiv, Jul. 2016. DOI: 10.48550/arXiv.1511.04599. [Online]. Available: http...

    2016 doi arXiv

  8. [16]

    The Space of Adversarial Strategies,

    R. Sheatsley, B. Hoak, E. Pauley, and P. McDaniel, “The Space of Adversarial Strategies,” en, in USENIX Security 2023, Number: arXiv:2209.04521 arXiv:2209.04521 [cs], arXiv, Sep. 2022. [Online]. Available: http://arxiv.org/ abs/2209.04521

    2023 arXiv

  9. [17]

    Towards Evaluating the Robustness of Neural Networks,

    N. Carlini and D. Wagner, “Towards Evaluating the Robustness of Neural Networks,” in IEEE S&P 2017 , arXiv, Mar. 2017. DOI: 10.48550/arXiv.1608.04644. [Online]. Available: http://arxiv.org/abs/1608.04644

    2017 doi arXiv

  10. [18]

    The Limitations of Deep Learning in Adversarial Settings,

    N. Papernot, P. McDaniel, S. Jha, M. Fredrikson, Z. B. Celik, and A. Swami, “The Limitations of Deep Learning in Adversarial Settings,” in IEEE Euro S&P 2016, arXiv, Nov. 2015. DOI: 10.48550/arXiv.1511.07528. [Online]. Available: http://arxiv.org/abs/1511.07528

    2016 doi arXiv

  11. [19]

    ImageNet Large Scale Visual Recognition Challenge,

    O. Russakovsky, J. Deng, H. Su, et al., “ImageNet Large Scale Visual Recognition Challenge,” en, in IJCV 2015, arXiv, Jan. 2015. [Online]. Available: http://arxiv.org/ abs/1409.0575

    2015 arXiv

  12. [20]

    Torchvision the machine- vision package of torch,

    S. Marcel and Y . Rodriguez, “Torchvision the machine- vision package of torch,” in Proceedings of the 18th ACM international conference on Multimedia , ser. MM ’10, New York, NY , USA: Association for Computing Machinery, Oct. 2010, pp. 1485–1488, ISBN : 978-1- 60558-933-6. D...

    2010

  13. [21]

    Deep Residual Learning for Image Recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,” in IEEE/CVF Con- ference on Computer Vision and Pattern Recognition (CVPR) 2016, arXiv, Dec. 2015. DOI: 10.48550/arXiv. 1512.03385. [Online]. Available: http://arxiv.org/abs/ 1512.03385

    2016 doi arXiv

  14. [22]

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,

    M. Tan and Q. V . Le, “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,” in Interna- tional Conference on Machine Learning (ICML) 2019 , arXiv, Sep. 2020. DOI: 10.48550/arXiv.1905.11946. [Online]. Available: http://arxiv.org/abs/1905.11946

    2019 doi arXiv

  15. [23]

    Densely Connected Convolutional Networks,

    G. Huang, Z. Liu, L. van der Maaten, and K. Q. Wein- berger, “Densely Connected Convolutional Networks,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2017 , arXiv, Jan. 2018. DOI: 10. 48550 / arXiv. 1608 . 06993. [Online]. Available: http : //arxiv....

    2017 doi arXiv

  16. [24]

    Rethinking the Inception Architecture for Computer Vision,

    C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wo- jna, “Rethinking the Inception Architecture for Computer Vision,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2016, arXiv, Dec. 2015. DOI: 10.48550/arXiv.1512.00567. [Online]. Available: ht...

    2016 doi arXiv

  17. [25]

    Available: http://arxiv.org/abs/1708.06131

    [Online]. Available: http://arxiv.org/abs/1708.06131

    arXiv

  18. [26]

    The Origins and Prevalence of Texture Bias in Convolutional Neural Networks,

    K. L. Hermann, T. Chen, and S. Kornblith, “The Origins and Prevalence of Texture Bias in Convolutional Neural Networks,” en, in Conference on Neural Information Processing Systems (NeurIPS) 2020 , arXiv, Nov. 2020. [Online]. Available: http://arxiv.org/abs/1911.09071

    2020 arXiv

  19. [27]

    A ConvNet for the 2020s,

    Z. Liu, H. Mao, C. -Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,” en, in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2022 , Jan. 2022. [Online]. Available: https:// arxiv.org/abs/2201.03545v2

    2022 arXiv

  20. [28]

    Texture Synthesis Using Convolutional Neural Networks,

    L. A. Gatys, A. S. Ecker, and M. Bethge, “Texture Synthesis Using Convolutional Neural Networks,” en, in Conference on Neural Information Processing Systems (NeurIPS) 2015, arXiv, Nov. 2015. [Online]. Available: http://arxiv.org/abs/1505.07376

    2015 arXiv

  21. [29]

    Network Dissection: Quantifying Interpretability of Deep Visual Representations,

    D. Bau, B. Zhou, A. Khosla, A. Oliva, and A. Torralba, “Network Dissection: Quantifying Interpretability of Deep Visual Representations,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2017, arXiv, Apr. 2017. DOI: 10.48550/arXiv.1704. 05796. [Online]....

    2017 doi

  22. [30]

    Benchmarking Neu- ral Network Robustness to Common Corruptions and Perturbations,

    D. Hendrycks and T. Dietterich, “Benchmarking Neu- ral Network Robustness to Common Corruptions and Perturbations,” in International Conference on Learning Representations (ICLR) 2019 , arXiv, Mar. 2019. DOI: 10.48550/arXiv.1903.12261. [Online]. Available: http: //arxiv.org/ab...

    2019 doi arXiv

  23. [31]

    Image Style Transfer Using Convolutional Neural Networks,

    L. A. Gatys, A. S. Ecker, and M. Bethge, “Image Style Transfer Using Convolutional Neural Networks,” en, in 2016 IEEE Conference on Computer Vision and Pattern Recognition (IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)), Las Vegas, NV , USA: IEEE, Jun. ...

    2016

  24. [32]

    Shape-Texture Debiased Neural Network Training,

    Y . Li, Q. Yu, M. Tan, et al., “Shape-Texture Debiased Neural Network Training,” en, in International Confer- ence on Learning Representations (ICLR) 2021 , arXiv, Mar. 2021. [Online]. Available: http://arxiv.org/abs/2010. 05981

    2021

  25. [33]

    Shift from Texture-bias to Shape-bias: Edge Deformation-based Augmentation for Robust Object Recognition,

    X. He, Q. Lin, C. Luo, et al., “Shift from Texture-bias to Shape-bias: Edge Deformation-based Augmentation for Robust Object Recognition,” en, in 2023 IEEE/CVF International Conference on Computer Vision (ICCV) , Paris, France: IEEE, Oct. 2023, pp. 1526–1535. DOI: 10.1109/ICCV...

    2023

  26. [34]

    Adversarial Machine Learning at Scale,

    A. Kurakin, I. Goodfellow, and S. Bengio, “Adversarial Machine Learning at Scale,” in International Conference on Learning Representations (ICLR) 2017 , arXiv, Feb

    2017

  27. [35]

    Huang, B

    R. Huang, B. Xu, D. Schuurmans, and C. Szepesvari, Learning with a Strong Adversary , Jan. 2016. DOI: 10. 48550 / arXiv. 1511 . 03034. [Online]. Available: http : //arxiv.org/abs/1511.03034

    2016 doi arXiv

  28. [36]

    L. A. Gatys, A. S. Ecker, and M. Bethge, A Neural Algorithm of Artistic Style , en, Sep. 2015. [Online]. Available: http://arxiv.org/abs/1508.06576

    2015 arXiv

  29. [37]

    Deep Learning based Feature Ex- traction for Texture Classification,

    P. Simon and U. V, “Deep Learning based Feature Ex- traction for Texture Classification,” Procedia Computer Science, Third International Conference on Computing and Network Communications (CoCoNet’19), vol. 171, pp. 1680–1687, Jan. 2020, ISSN : 1877-0509. DOI: 10. 1016 / j . p...

    2020

  30. [38]

    Interpreting Adversarially Trained Convolutional Neural Networks,

    T. Zhang and Z. Zhu, “Interpreting Adversarially Trained Convolutional Neural Networks,” en, in International Conference on Machine Learning (ICML) 2019 , arXiv, May 2019. [Online]. Available: http://arxiv.org/abs/1905. 09797

    2019

  31. [39]

    Explore the Transfor- mation Space for Adversarial Images,

    J. Chen, D. Wang, and H. Chen, “Explore the Transfor- mation Space for Adversarial Images,” in Proceedings of the Tenth ACM Conference on Data and Application Security and Privacy , ser. CODASPY ’20, New York, NY , USA: Association for Computing Machinery, Mar. 2020, pp. 109–1...

    2020

  32. [40]

    Color Channel Perturbation Attacks for Fooling Convolutional Neural Networks and A Defense Against Such Attacks,

    J. Kantipudi, S. R. Dubey, and S. Chakraborty, “Color Channel Perturbation Attacks for Fooling Convolutional Neural Networks and A Defense Against Such Attacks,” in IEEE Transactions on Artificial Intelligence , arXiv, Dec. 2020. DOI: 10.48550/arXiv.2012.14456. [Online]. Avail...

    2020 doi arXiv

  33. [41]

    Color encoding in biologically-inspired convolutional neural networks,

    I. Rafegas and M. Vanrell, “Color encoding in biologically-inspired convolutional neural networks,” Vision Research, Color: cone opponency and beyond, vol. 151, pp. 7–17, Oct. 2018, ISSN : 0042-6989. DOI: 10 . 1016 / j . visres . 2018 . 03 . 010. [Online]. Available: https : /...

    2018

  34. [44]

    Universal adversarial perturbations,

    S.-M. Moosavi-Dezfooli, A. Fawzi, O. Fawzi, and P. Frossard, “Universal adversarial perturbations,” in IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2017 , arXiv, Mar. 2017. DOI: 10. 48550 / arXiv. 1610 . 08401. [Online]. Available: http : //arxiv.org/a...

    2017 doi arXiv

  35. [45]

    These files contain the necessary instructions, images, and the script you will run for this study

    Download the Package Download the provided tarball package of your choosing, the eval_packages.py script, and the README_humaneval.md file. These files contain the necessary instructions, images, and the script you will run for this study. Once you have downloaded everything, ...

  36. [46]

    python3 eval_packages.py package_num The script will display 100 images, one at a time in a pop-up window along with four words in the terminal

    Run the Script Open a terminal in the directory where you placed the files and run the Python script with {package_num} being the package number (shown in the tarball name) you would like to evaluate. python3 eval_packages.py package_num The script will display 100 images, one...

  37. [47]

    Your task is to input the number corre- sponding to the texture that you believe is most prominent in the image

    Input Your Responses For each image, you will be shown four texture words. Your task is to input the number corre- sponding to the texture that you believe is most prominent in the image. Note that you do not have to click out of the current texture image; inputting your answe...

  38. [48]

    Complete the Study Once you have completed the evaluation for all 100 images, a completion message will show, and the script will save your results in a CSV file

  39. [49]

    Upload Your Results Please upload the generated CSV file to the provided Google Drive link. B. Additional TID examples Figure 15, Figure 16, Figure 17, Figure 18, and Figure 19 show examples of ImageNet validation images identified by the TID of ResNet50 as having various text...

  40. [2017]

    48550 / arXiv

    DOI: 10 . 48550 / arXiv . 1611 . 01236. [Online]. Available: http://arxiv.org/abs/1611.01236

    doi arXiv

  41. [2019]

    Available: http://arxiv.org/abs/1811

    [Online]. Available: http://arxiv.org/abs/1811. 12231

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.