Pith. sign in

REVIEW 3 major objections 7 minor 64 references

Stylized cue-conflict images give unstable shape-texture bias scores; pure, balanced cues plus full-label ranking fix the diagnosis.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-14 23:17 UTC pith:2OC2LEIN

load-bearing objection Solid empirical fix for a widely used but noisy diagnostic: cleaner pure-cue stimuli + full-label MRR sensitivity that actually recovers expected training effects and resolves the shape-vs-texture contradictions in the cue-conflict literature. the 3 major comments →

arxiv 2603.10834 v4 pith:2OC2LEIN submitted 2026-03-11 cs.CV cs.AI

On the Reliability of Cue Conflict and Beyond

Pum Jun Kim , Seung-Ah Lee , Seongho Park , Dongyoon Han , Jaejun Yoo This is my paper
classification cs.CV cs.AI
keywords shape-texture biascue conflictneural network interpretabilityvisual cuesbias evaluationmean reciprocal rankcomputer vision
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Neural networks are often judged by how much they prefer object shape over surface texture, and the standard test has long been stylized images that mash one class's shape with another's texture. The paper shows that this test is unreliable: stylization leaks cues into each other, leaves them unbalanced and hard to recognize, ratio scores hide absolute sensitivity, and evaluating only a handful of labels warps the model's real predictions. The authors therefore build REFINED-BIAS: clean shape and texture images defined by human-perceptible geometry and repeating patterns, plus a ranking metric that scores how highly each true cue sits in the full output list. With this setup, training strategies that should increase shape use actually do so, both cues together track accuracy, and earlier contradictory claims about which cue helps performance resolve into a consistent picture. Readers who care about trustworthy, human-interpretable diagnostics for vision models now have a cleaner instrument.

Core claim

The stylization-based cue-conflict benchmark confounds genuine preference with invalid, unbalanced, or unrecognizable cues and with evaluation artifacts; REFINED-BIAS restores reliable diagnosis by constructing pure, balanced, human- and model-recognizable shape and texture stimuli under explicit perceptual definitions and by measuring absolute cue sensitivity over the full label space with a ranking metric.

What carries the argument

REFINED-BIAS: a curated set of pure shape (global+local geometry via segmentation, blur, and edges) and pure texture (scale-consistent interior patches reordered to remove structure) images, scored by mean reciprocal rank of the correct shape or texture label in the model's full prediction list (Shape-Sens and Texture-Sens), from which preference ratios can still be derived.

Load-bearing premise

The authors' hand-curated definitions of pure shape and pure texture, plus the generation pipeline, actually produce cues that are separable, equally informative, free of residual confounds, and free of domain shift for the networks under test.

What would settle it

If human and model recognition accuracies on the new pure shape and pure texture sets remain heavily imbalanced or fail to correlate with independent depth-map and artifact-free texture proxies, or if shape-focused training strategies still show no reliable rise in Shape-Sens under the ranking metric, the claim that REFINED-BIAS removes the confounding artifacts collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Bias comparisons across architectures and training recipes become fairer because absolute sensitivity is reported alongside preference.
  • Shape-focused augmentations and local-to-global designs (shifted windows, conv-attention hybrids) can be shown to raise shape sensitivity rather than merely changing ratios.
  • In-domain accuracy tracks joint use of both cues once the confounds are removed, replacing earlier contradictory shape-versus-texture conclusions.
  • Future bias benchmarks can reuse the pure-cue construction and full-label ranking protocol instead of stylization and restricted labels.

Where Pith is reading between the lines

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

  • Any diagnostic that still mixes two cues inside a single image will re-introduce the same validity and balance problems unless it also reports separate pure-cue sensitivities.
  • The ranking metric could be ported to other multi-cue settings (color versus form, motion versus form) where accuracy ratios currently hide absolute utilization.
  • If residual 3-D or viewpoint information remains in the shape silhouettes, later viewpoint-controlled tests will be needed to isolate purely 2-D geometric bias.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 7 minor

Summary. The paper argues that the widely used stylization-based cue-conflict benchmark yields unstable shape–texture bias estimates because stylization imperfectly separates and balances cues, ratio-based scores obscure absolute sensitivity, and restricted-label evaluation distorts predictions. It introduces REFINED-BIAS: a curated set of pure shape (segmentation + blur + edges) and texture (interior patches reordered to remove structure) stimuli over 20 ImageNet-derived superclasses, plus full-label ranking metrics (Shape-Sens / Texture-Sens via mean reciprocal rank, and their ratio as preference). Across ResNet-50 training strategies and multiple CNN/ViT architectures, the framework is claimed to recover expected effects of shape-focused training, separate sensitivity from preference, resolve prior contradictory conclusions about shape vs. texture and in-domain accuracy, and better match human recognizability (Fleiss’ κ, top-1 rates) than cue-conflict and two recent alternatives.

Significance. If the empirical claims hold, the work supplies a more reliable diagnostic for a de facto standard in vision bias analysis and would clarify conflicting literature on whether shape or texture preference drives ImageNet performance. Strengths include public code, parallel human studies with inter-rater agreement, confidence intervals, t-tests against training-strategy hypotheses, side-by-side comparison with cue-conflict and with Tartaglini et al. and Burgert et al., and proxy checks (depth maps; grid-free textures). The contribution is primarily methodological and empirical rather than theoretical; its lasting value depends on community adoption of the released stimuli and on whether pure-cue sensitivity generalizes as a stand-in for preference under true cue competition.

major comments (3)
  1. §3.1–3.2 and §5: Preference is defined from separate pure-cue Shape-Sens and Texture-Sens (Eqs. 1–3), not from simultaneous conflict stimuli. This is a paradigm shift relative to cue-conflict. The Discussion asserts that pure-cue sensitivity yields a more grounded preference, but the manuscript does not show that relative pure-cue ranks predict behavior when both cues are present in the same image (e.g., by constructing controlled conflict pairs from the same pure sources and correlating the two preference scores). Without that check, claims that REFINED-BIAS measures “preference” (as opposed to absolute cue sensitivity) rest on an untested identification assumption.
  2. §3.1, Appendix A.1–A.2: Class selection (10 shape-dominant + 10 texture-dominant), patch sizes, blur/edge parameters, and manual filtering are free design choices that define the benchmark. The paper reports strong human κ and higher model top-1 than cue-conflict, which supports improved recognizability, but does not quantify sensitivity of the main conclusions (Table 2; Figs. 8–9) to alternative class sets or to the rejected samples. A leave-class-out or alternative-superclass ablation, or release of the pre-curation pool and rejection criteria, would make the head-to-head superiority claims more robust to the curation risk the authors themselves flag as a limitation.
  3. §4.1–4.2, Figs. 8–9: Architecture comparisons mix families trained with different default augmentations (CNNs with random crop; ViTs with mixed aug), while strategy comparisons fix ResNet-50. The paper notes family-wise comparison in Appendix B, but the main-text narrative sometimes treats architecture and strategy effects as cleanly separable. Explicitly stating which panels are confounded by training recipe, and optionally re-running a subset of architectures under a matched recipe, would strengthen the local-to-global / RF-size interpretation in Appendix A.11.
minor comments (7)
  1. §5 Discussion: the paragraph “What exactly is improved by the dataset, the metric, and their combination?” is duplicated verbatim; remove the second copy.
  2. Fig. 1–2 and several appendix figures use dense multi-panel layouts with small text; ensure axis labels and legends remain legible in print and that color encodings (shape vs texture) are colorblind-safe.
  3. Table 2 and Table A report p-values with shading for significance; also report effect sizes (e.g., mean preference difference and CI) so readers can judge practical magnitude, not only null-hypothesis rejection.
  4. §2.2.1 / Fig. 3b: Kendall correlations of stylized content/style with silhouette/edge proxies are informative; state the number of models and whether correlations are computed on class-wise accuracies pooled or per-model then averaged.
  5. Appendix A.3: human study n = 66 (REFINED-BIAS) vs 22 (cue-conflict); note the imbalance when comparing κ and accuracy CIs, or equalize sample sizes in a revision.
  6. Clarify in §3.2 that MRR is computed over the full ImageNet-1k (or model) label space and that ranks are 1-indexed; a short pseudocode block would aid reimplementation.
  7. Typos / polish: “REFINED-BIAS” spacing and capitalization vary slightly in places; “In-domain Acc” axis labels; ensure arXiv link and code URL are final in the camera-ready abstract.

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark whose metric and stimuli are design choices applied to external models, not results forced by construction or self-citation.

full rationale

The paper is an empirical diagnostic framework, not a first-principles derivation. Shape and texture are operationally defined (global/local non-repeating geometry vs. scale-consistent repetitive interior patterns), stimuli are constructed via segmentation/blur/edge extraction and patch reordering plus manual curation, and sensitivity is measured by mean reciprocal rank of the ground-truth labels in the full logit ranking (Eqs. 1–3). These are independent design choices; the subsequent rankings of training strategies and architectures (Tables 2, Figs. 8–9) are measurements on held-out pretrained models, not quantities fitted to produce the desired ordering. Human Fleiss’ κ, model top-1 accuracies, and Kendall correlations against depth/texture proxies serve as external checks rather than circular inputs. No uniqueness theorem, ansatz, or load-bearing self-citation forces the conclusions; residual risks (class selection, curation) are acknowledged design limitations, not definitional reductions. Score 0 is therefore warranted.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 1 invented entities

The central empirical claims rest on a small set of domain definitions of "shape" and "texture", a hand-chosen set of 20 superclasses, and the assumption that the extraction pipeline plus human filtering yields pure balanced cues. No free parameters are fitted to force the main performance-bias correlations; the free choices are design decisions for the benchmark itself.

free parameters (3)
  • number of superclasses (10 shape-dominant + 10 texture-dominant)
    Chosen by the authors for balance and recognizability; different selections could alter measured sensitivities.
  • patch sizes and reordering strategy for texture cues
    Four predefined sizes and a quilting-like reorder are fixed by design; they control residual structure leakage.
  • Gaussian blur strength and edge rendering for shape cues
    Class-adjusted blur and white-on-black edges are hand-tuned parameters of the generation pipeline.
axioms (3)
  • domain assumption Shape is non-repeating geometric structure (global silhouette + local distinctive substructures); texture is scale-consistent repetitive patterns.
    Stated in §3.1 as the human-perceptual definition used to construct stimuli; not derived from data.
  • ad hoc to paper Full-label ranking (MRR of the true shape/texture class) is a more faithful measure of cue sensitivity than restricted-label accuracy ratios.
    Introduced in §3.2 to address Problems 2 and 3; justified by argument rather than external theorem.
  • domain assumption ImageNet-1k pretrained models' top-1 / ranking behavior on the constructed cues reflects genuine shape/texture utilization rather than residual low-level artifacts.
    Implicit throughout the experimental sections; supported by higher absolute accuracies and human agreement but not proven free of confounds.
invented entities (1)
  • REFINED-BIAS dataset and Shape-Sens / Texture-Sens metrics no independent evidence
    purpose: Provide pure balanced stimuli and absolute cue-sensitivity scores that replace stylized cue-conflict ratios.
    New public artifacts introduced by the paper; independent evidence will come from third-party use and re-measurement.

pith-pipeline@v1.1.0-grok45 · 27753 in / 2670 out tokens · 25424 ms · 2026-07-14T23:17:22.673032+00:00 · methodology

0 comments
read the original abstract

Understanding how neural networks rely on visual cues offers a human-interpretable view of their internal decision processes. The cue-conflict benchmark has been influential in probing shape-texture preference and in motivating the insight that stronger, human-like shape bias is often associated with improved in-domain performance. However, we find that the current stylization-based instantiation can yield unstable and ambiguous bias estimates. Specifically, stylization may not reliably instantiate perceptually valid and separable cues nor control their relative informativeness, ratio-based bias can obscure absolute cue sensitivity, and restricting evaluation to preselected classes can distort model predictions by ignoring the full decision space. Together, these factors can confound preference with cue validity, cue balance, and recognizability artifacts. We introduce REFINED-BIAS, an integrated dataset and evaluation framework for reliable and interpretable shape-texture bias diagnosis. REFINED-BIAS constructs balanced, human- and model- recognizable cue pairs using explicit definitions of shape and texture, and measures cue-specific sensitivity over the full label space via a ranking-based metric, enabling fairer cross-model comparisons. Across diverse training regimes and architectures, REFINED-BIAS enables fairer cross-model comparison, more faithful diagnosis of shape and texture biases, and clearer empirical conclusions, resolving inconsistencies that prior cue-conflict evaluations could not reliably disambiguate.

discussion (0)

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

Reference graph

Works this paper leans on

64 extracted references · 15 linked inside Pith

  1. [1]

    Deep convolutional networks do not classify based on global object shape.PLoS computational biology, 14(12): e1006613, 2018

    Nicholas Baker, Hongjing Lu, Gennady Erlikhman, and Philip J Kellman. Deep convolutional networks do not classify based on global object shape.PLoS computational biology, 14(12): e1006613, 2018

    2018

  2. [2]

    The origins and prevalence of texture bias in convolutional neural networks.Advances in Neural Information Processing Systems, 33: 19000–19015, 2020

    Katherine Hermann, Ting Chen, and Simon Kornblith. The origins and prevalence of texture bias in convolutional neural networks.Advances in Neural Information Processing Systems, 33: 19000–19015, 2020

    2020

  3. [3]

    Partial success in closing the gap between human and machine vision.Advances in Neural Information Processing Systems, 34:23885– 23899, 2021

    Robert Geirhos, Kantharaju Narayanappa, Benjamin Mitzkus, Tizian Thieringer, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Partial success in closing the gap between human and machine vision.Advances in Neural Information Processing Systems, 34:23885– 23899, 2021

    2021

  4. [4]

    Scaling up your kernels to 31x31: Revisiting large kernel design in cnns

    Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11963–11975, 2022

    2022

  5. [5]

    An extended study of human-like behavior under adversarial training

    Paul Gavrikov, Janis Keuper, and Margret Keuper. An extended study of human-like behavior under adversarial training. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2361–2368, 2023

    2023

  6. [6]

    Wavelet convolutions for large receptive fields

    Shahaf E Finder, Roy Amoyal, Eran Treister, and Oren Freifeld. Wavelet convolutions for large receptive fields. InEuropean Conference on Computer Vision, pages 363–380. Springer, 2024

    2024

  7. [7]

    Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. InInternational Conference on Learning Representations, 2018

    2018

  8. [8]

    What shapes feature representations? exploring datasets, architectures, and training.Advances in Neural Information Processing Systems, 33: 9995–10006, 2020

    Katherine Hermann and Andrew Lampinen. What shapes feature representations? exploring datasets, architectures, and training.Advances in Neural Information Processing Systems, 33: 9995–10006, 2020

    2020

  9. [9]

    Teaching cnns to mimic human visual cognitive process & regularise texture-shape bias

    Satyam Mohla, Anshul Nasery, and Biplab Banerjee. Teaching cnns to mimic human visual cognitive process & regularise texture-shape bias. InICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1805–1809. IEEE, 2022

    2022

  10. [10]

    Does resistance to style-transfer equal global shape bias? measuring network sensitivity to global shape configuration.arXiv preprint arXiv:2310.07555, 2023

    Ziqi Wen, Tianqin Li, Zhi Jing, and Tai Sing Lee. Does resistance to style-transfer equal global shape bias? measuring network sensitivity to global shape configuration.arXiv preprint arXiv:2310.07555, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Visual anagrams reveal hidden differences in holistic shape processing across vision models.arXiv preprint arXiv:2507.00493, 2025

    Fenil R Doshi, Thomas Fel, Talia Konkle, and George Alvarez. Visual anagrams reveal hidden differences in holistic shape processing across vision models.arXiv preprint arXiv:2507.00493, 2025

    arXiv 2025

  12. [12]

    Imagenet-trained cnns are not biased towards texture: Revisiting feature reliance through controlled suppression.arXiv preprint arXiv:2509.20234, 2025

    Tom Burgert, Oliver Stoll, Paolo Rota, and Begüm Demir. Imagenet-trained cnns are not biased towards texture: Revisiting feature reliance through controlled suppression.arXiv preprint arXiv:2509.20234, 2025

    arXiv 2025

  13. [13]

    Interpreting adversarially trained convolutional neural networks

    Tianyuan Zhang and Zhanxing Zhu. Interpreting adversarially trained convolutional neural networks. InInternational conference on machine learning, pages 7502–7511. PMLR, 2019

    2019

  14. [14]

    Does enhanced shape bias improve neural network robustness to common corruptions?arXiv preprint arXiv:2104.09789, 2021

    Chaithanya Kumar Mummadi, Ranjitha Subramaniam, Robin Hutmacher, Julien Vitay, V olker Fischer, and Jan Hendrik Metzen. Does enhanced shape bias improve neural network robustness to common corruptions?arXiv preprint arXiv:2104.09789, 2021

    Pith/arXiv arXiv 2021

  15. [15]

    Connecting metrics for shape-texture knowledge in computer vision.arXiv preprint arXiv:2301.10608, 2023

    Tiago Oliveira, Tiago Marques, and Arlindo L Oliveira. Connecting metrics for shape-texture knowledge in computer vision.arXiv preprint arXiv:2301.10608, 2023

    Pith/arXiv arXiv 2023

  16. [16]

    Emergence of shape bias in convolutional neural networks through activation sparsity.Advances in Neural Information Processing Systems, 36:71755–71766, 2023

    Tianqin Li, Ziqi Wen, Yangfan Li, and Tai Sing Lee. Emergence of shape bias in convolutional neural networks through activation sparsity.Advances in Neural Information Processing Systems, 36:71755–71766, 2023. 11

    2023

  17. [17]

    Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22184–22194, 2024

    Paul Gavrikov and Janis Keuper. Can biases in imagenet models explain generalization? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22184–22194, 2024

    2024

  18. [18]

    Toward a holistic evaluation of robustness in clip models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Weijie Tu, Weijian Deng, and Tom Gedeon. Toward a holistic evaluation of robustness in clip models.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  19. [19]

    Percep- tual inductive bias is what you need before contrastive learning

    Junru Zhao, Tianqin Li, Dunhan Jiang, Shenghao Wu, Alan Ramirez, and Tai Sing Lee. Percep- tual inductive bias is what you need before contrastive learning. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 9621–9630, 2025

    2025

  20. [20]

    On the dynamic evolution of clip texture-shape bias and its relationship to human alignment and model robustness.arXiv preprint arXiv:2508.09814, 2025

    Pablo Hernández-Cámara, Jose Manuel Jaén-Lorites, Alexandra Gómez-Villa, Jorge Vila-Tomás, Valero Laparra, and Jesus Malo. On the dynamic evolution of clip texture-shape bias and its relationship to human alignment and model robustness.arXiv preprint arXiv:2508.09814, 2025

    arXiv 2025

  21. [21]

    L1-induced activation sparsity shifts cnn bias from texture to shape

    Sungjae Jeon, Youjin Kim, Seungtae Hong, and Jeong-Si Kim. L1-induced activation sparsity shifts cnn bias from texture to shape. In2025 IEEE/IEIE International Conference on Consumer Electronics-Asia (ICCE-Asia), pages 1–4. IEEE, 2025

    2025

  22. [22]

    Enhanc- ing 2d representation learning with a 3d prior

    Mehmet Aygun, Prithviraj Dhar, Zhicheng Yan, Oisin Mac Aodha, and Rakesh Ranjan. Enhanc- ing 2d representation learning with a 3d prior. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7750–7760, 2024

    2024

  23. [23]

    Contour integration underlies human-like vision.arXiv preprint arXiv:2504.05253, 2025

    Ben Lonnqvist, Elsa Scialom, Abdulkadir Gokce, Zehra Merchant, Michael H Herzog, and Mar- tin Schrimpf. Contour integration underlies human-like vision.arXiv preprint arXiv:2504.05253, 2025

    Pith/arXiv arXiv 2025

  24. [24]

    Contributions of shape, texture, and color in visual recognition

    Yunhao Ge, Yao Xiao, Zhi Xu, Xingrui Wang, and Laurent Itti. Contributions of shape, texture, and color in visual recognition. InEuropean Conference on Computer Vision, pages 369–386. Springer, 2022

    2022

  25. [25]

    Image style transfer using convolutional neural networks

    Leon A Gatys, Alexander S Ecker, and Matthias Bethge. Image style transfer using convolutional neural networks. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2414–2423, 2016

    2016

  26. [26]

    Generalisation in humans and deep neural networks.Advances in Neural Information Processing Systems, 31, 2018

    Robert Geirhos, Carlos RM Temme, Jonas Rauber, Heiko H Schütt, Matthias Bethge, and Felix A Wichmann. Generalisation in humans and deep neural networks.Advances in Neural Information Processing Systems, 31, 2018

    2018

  27. [27]

    A developmentally-inspired examination of shape versus texture bias in machines.arXiv preprint arXiv:2202.08340, 2022

    Alexa R Tartaglini, Wai Keen V ong, and Brenden M Lake. A developmentally-inspired examination of shape versus texture bias in machines.arXiv preprint arXiv:2202.08340, 2022

    Pith/arXiv arXiv 2022

  28. [28]

    Measuring nominal scale agreement among many raters.Psychological bulletin, 76(5):378, 1971

    Joseph L Fleiss. Measuring nominal scale agreement among many raters.Psychological bulletin, 76(5):378, 1971

    1971

  29. [29]

    The trec-8 question answering track report

    Ellen M V oorhees et al. The trec-8 question answering track report. InTrec, volume 99, pages 77–82, 1999

    1999

  30. [30]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016

    2016

  31. [31]

    Shape-texture debiased neural network training.arXiv preprint arXiv:2010.05981, 2020

    Yingwei Li, Qihang Yu, Mingxing Tan, Jieru Mei, Peng Tang, Wei Shen, Alan Yuille, and Cihang Xie. Shape-texture debiased neural network training.arXiv preprint arXiv:2010.05981, 2020

    Pith/arXiv arXiv 2010

  32. [32]

    An empirical study of training self-supervised vision transformers

    Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9640–9649, 2021

    2021

  33. [33]

    Big self-supervised models are strong semi-supervised learners.Advances in Neural Information Processing Systems, 33:22243–22255, 2020

    Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey E Hinton. Big self-supervised models are strong semi-supervised learners.Advances in Neural Information Processing Systems, 33:22243–22255, 2020. 12

    2020

  34. [34]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9650–9660, 2021

    2021

  35. [35]

    Augmix: A simple data processing method to improve robustness and uncertainty

    Dan Hendrycks, Norman Mu, Ekin D Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshmi- narayanan. Augmix: A simple data processing method to improve robustness and uncertainty. arXiv preprint arXiv:1912.02781, 2019

    Pith/arXiv arXiv 1912

  36. [36]

    Pixmix: Dreamlike pictures comprehensively improve safety measures

    Dan Hendrycks, Andy Zou, Mantas Mazeika, Leonard Tang, Bo Li, Dawn Song, and Jacob Steinhardt. Pixmix: Dreamlike pictures comprehensively improve safety measures. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16783–16792, 2022

    2022

  37. [37]

    Prime: A few primitives can boost robustness to common corruptions

    Apostolos Modas, Rahul Rade, Guillermo Ortiz-Jiménez, Seyed-Mohsen Moosavi-Dezfooli, and Pascal Frossard. Prime: A few primitives can boost robustness to common corruptions. In European Conference on Computer Vision, pages 623–640. Springer, 2022

    2022

  38. [38]

    Classification robustness to common optical aberrations

    Patrick Müller, Alexander Braun, and Margret Keuper. Classification robustness to common optical aberrations. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3632–3643, 2023

    2023

  39. [39]

    The many faces of robustness: A critical analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340–8349, 2021

    2021

  40. [40]

    Resnet strikes back: An improved training procedure in timm.arXiv preprint arXiv:2110.00476, 2021

    Ross Wightman, Hugo Touvron, and Hervé Jégou. Resnet strikes back: An improved training procedure in timm.arXiv preprint arXiv:2110.00476, 2021

    Pith/arXiv arXiv 2021

  41. [41]

    Randaugment: Practical automated data augmentation with a reduced search space

    Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 702–703, 2020

    2020

  42. [42]

    Do ad- versarially robust imagenet models transfer better?Advances in Neural Information Processing Systems, 33:3533–3545, 2020

    Hadi Salman, Andrew Ilyas, Logan Engstrom, Ashish Kapoor, and Aleksander Madry. Do ad- versarially robust imagenet models transfer better?Advances in Neural Information Processing Systems, 33:3533–3545, 2020

    2020

  43. [43]

    Enhancing shape bias for object detection.Neurocomputing, page 132931, 2026

    Jiwen Tang, Gu Wang, Ruida Zhang, and Xiangyang Ji. Enhancing shape bias for object detection.Neurocomputing, page 132931, 2026

    2026

  44. [44]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Pith/arXiv arXiv 2010

  45. [45]

    Tokens-to-token vit: Training vision transformers from scratch on imagenet

    Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zi-Hang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch on imagenet. InProceedings of the IEEE/CVF international conference on computer vision, pages 558–567, 2021

    2021

  46. [46]

    Conformer: Local features coupling global representations for visual recognition

    Zhiliang Peng, Wei Huang, Shanzhi Gu, Lingxi Xie, Yaowei Wang, Jianbin Jiao, and Qixiang Ye. Conformer: Local features coupling global representations for visual recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 367–376, 2021

    2021

  47. [47]

    Cvt: Introducing convolutions to vision transformers

    Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 22–31, 2021

    2021

  48. [48]

    Visualization comparison of vision transformers and convolutional neural networks.IEEE Transactions on Multimedia, 26: 2327–2339, 2023

    Rui Shi, Tianxing Li, Liguo Zhang, and Yasushi Yamaguchi. Visualization comparison of vision transformers and convolutional neural networks.IEEE Transactions on Multimedia, 26: 2327–2339, 2023. 13

    2023

  49. [49]

    Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159, 2024

    Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks.arXiv preprint arXiv:2401.14159, 2024

    Pith/arXiv arXiv 2024

  50. [50]

    Depth anything v2.Advances in Neural Information Processing Systems, 37:21875– 21911, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2.Advances in Neural Information Processing Systems, 37:21875– 21911, 2024

    2024

  51. [51]

    Image quilting for texture synthesis and transfer

    Alexei A Efros and William T Freeman. Image quilting for texture synthesis and transfer. In Seminal graphics papers: pushing the boundaries, volume 2, pages 571–576, 2023

    2023

  52. [52]

    Computing receptive fields of convolutional neural networks.Distill, 2019

    André Araujo, Wade Norris, and Jack Sim. Computing receptive fields of convolutional neural networks.Distill, 2019. doi: 10.23915/distill.00021. https://distill.pub/2019/computing- receptive-fields

    doi:10.23915/distill.00021 2019

  53. [53]

    More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity.arXiv preprint arXiv:2207.03620, 2022

    Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity.arXiv preprint arXiv:2207.03620, 2022

    Pith/arXiv arXiv 2022

  54. [54]

    Scaling up your kernels: Large kernel design in convnets towards universal representations.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    Yiyuan Zhang, Xiaohan Ding, and Xiangyu Yue. Scaling up your kernels: Large kernel design in convnets towards universal representations.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

    2025

  55. [55]

    Large receptive field strategy and important feature extraction strategy in 3d object detection.arXiv preprint arXiv:2401.11913, 2024

    Leichao Cui, Xiuxian Li, Min Meng, and Guangyu Jia. Large receptive field strategy and important feature extraction strategy in 3d object detection.arXiv preprint arXiv:2401.11913, 2024

    Pith/arXiv arXiv 2024

  56. [56]

    Spikingyolox: Improved yolox object detection with fast fourier convolution and spiking neural networks

    Wei Miao, Jiangrong Shen, Qi Xu, Timo Hamalainen, Yi Xu, and Fengyu Cong. Spikingyolox: Improved yolox object detection with fast fourier convolution and spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 1465–1473, 2025

    2025

  57. [57]

    Approximating cnns with bag-of-local-features models works surprisingly well on imagenet.arXiv preprint arXiv:1904.00760, 2019

    Wieland Brendel and Matthias Bethge. Approximating cnns with bag-of-local-features models works surprisingly well on imagenet.arXiv preprint arXiv:1904.00760, 2019

    Pith/arXiv arXiv 1904

  58. [58]

    Imagenet classification with deep convolutional neural networks.Advances in Neural Information Processing Systems, 25, 2012

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks.Advances in Neural Information Processing Systems, 25, 2012

    2012

  59. [59]

    Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

    Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition.arXiv preprint arXiv:1409.1556, 2014

    Pith/arXiv arXiv 2014

  60. [60]

    Re- thinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Re- thinking the inception architecture for computer vision. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2818–2826, 2016

    2016

  61. [61]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021

    2021

  62. [62]

    Cmt: Convolutional neural networks meet vision transformers

    Jianyuan Guo, Kai Han, Han Wu, Yehui Tang, Xinghao Chen, Yunhe Wang, and Chang Xu. Cmt: Convolutional neural networks meet vision transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12175–12185, 2022

    2022

  63. [63]

    boat” and “car

    Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks.arXiv preprint arXiv:1706.06083, 2017. 14 On the Reliability of Cue Conflict and Beyond Supplementary Material A Details onREFINED-BIAS A.1 Our Pipeline for Cue Generation To ensure each cue provides a wel...

    Pith/arXiv arXiv 2017

  64. [64]

    Vision transformer variants, including ViT [44], Swin Transformer [61], and CMT [62], additionally employ mixed augmentation strategies during training

    are trained with standard random resized cropping. Vision transformer variants, including ViT [44], Swin Transformer [61], and CMT [62], additionally employ mixed augmentation strategies during training. Due to these differences, comparisons in Fig. 8 and 9 are conducted within each model family. Table B: Effect of training strategies on shape/texture sen...

    2065