mlsanityData to 7 Oct 2026

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

arXiv 1811.12231 ↗First posted 29 Nov 2018cs.CV cs.AI cs.LGcited by 1,430

Not on the map. Why?

Cited by 40 in Diffusion models, but placed by its own abstract, which didn’t fit.

Abstract (arXiv)

Convolutional Neural Networks (CNNs) are commonly thought to recognise objects by learning increasingly complex representations of object shapes. Some recent studies suggest a more important role of image textures. We here put these conflicting hypotheses to a quantitative test by evaluating CNNs and human observers on images with a texture-shape cue conflict. We show that ImageNet-trained CNNs are strongly biased towards recognising textures rather than shapes, which is in stark contrast to human behavioural evidence and reveals fundamentally different classification strategies. We then demonstrate that the same standard architecture (ResNet-50) that learns a texture-based representation on ImageNet is able to learn a shape-based representation instead when trained on "Stylized-ImageNet", a stylized version of ImageNet. This provides a much better fit for human behavioural performance in our well-controlled psychophysical lab setting (nine experiments totalling 48,560 psychophysical trials across 97 observers) and comes with a number of unexpected emergent benefits such as improved object detection performance and previously unseen robustness towards a wide range of image distortions, highlighting advantages of a shape-based representation.

Key facts. Cited by 1,430 papers in the map (data through 7 Oct 2026). First posted on arXiv 29 Nov 2018.

Builds on

No reference list yet.

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2021: 277; 2022: 278; 2023: 272; 2024: 245; 2025: 191; 2026: 167.

1,430 papers.

Sort:

By area:

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
  2. 2021Robust fine-tuning of zero-shot modelscited by 785
  3. 2021Resolution-robust Large Mask Inpainting with Fourier Convolutionscited by 730
  4. 2023Scaling Vision Transformers to 22 Billion Parameterscited by 690
  5. 2022When and why vision-language models behave like bags-of-words, and what to do about it?cited by 603
  6. 2022Mind the Gap: Understanding the Modality Gap in Multi-modal Contrastive Representation Learningcited by 553
  7. 2022Beyond neural scaling laws: beating power law scaling via data pruningcited by 519
  8. 2024Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Trainingcited by 478
  9. 2021Intriguing Properties of Vision Transformerscited by 446
  10. 2021SWAD: Domain Generalization by Seeking Flat Minimacited by 431
  11. 2023Directly Fine-Tuning Diffusion Models on Differentiable Rewardscited by 431
  12. 2021Revisiting the Calibration of Modern Neural Networkscited by 387
38 more
  1. 2022Last Layer Re-Training is Sufficient for Robustness to Spurious Correlationscited by 379
  2. 2023RemoteCLIP: A Vision Language Foundation Model for Remote Sensingcited by 357
  3. 2021Deep ViT Features as Dense Visual Descriptorscited by 327
  4. 2021Self-Supervised Learning with Data Augmentations Provably Isolates Content from Stylecited by 308
  5. 2021The Intrinsic Dimension of Images and Its Impact on Learningcited by 307
  6. 2022How Do Vision Transformers Work?cited by 305
  7. 2021DAFormer: Improving Network Architectures and Training Strategies for Domain-Adaptive Semantic Segmentationcited by 299
  8. 2021A Fourier-based Framework for Domain Generalizationcited by 287
  9. 2023Human Preference Score: Better Aligning Text-to-Image Models with Human Preferencecited by 284
  10. 2022CLIP-Mesh: Generating textured meshes from text using pretrained image-text modelscited by 281
  11. 2024Fast High-Resolution Image Synthesis with Latent Adversarial Diffusion Distillationcited by 276
  12. 2023TRAK: Attributing Model Behavior at Scalecited by 273
  13. 2021Generalizing Face Forgery Detection with High-frequency Featurescited by 261
  14. 2021Gradient Matching for Domain Generalizationcited by 258
  15. 2021Understanding Robustness of Transformers for Image Classificationcited by 250
  16. 2022Self-Supervised Contrastive Pre-Training For Time Series via Time-Frequency Consistencycited by 246
  17. 2022Delving into Out-of-Distribution Detection with Vision-Language Representationscited by 215
  18. 2022The Role of ImageNet Classes in Fr'echet Inception Distancecited by 211
  19. 2021Vision Transformers are Robust Learnerscited by 208
  20. 2025Video models are zero-shot learners and reasonerscited by 208
  21. 2021Towards Causal Representation Learningcited by 207
  22. 2023Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion modelscited by 199
  23. 2023Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfercited by 194
  24. 2021Are Transformers More Robust Than CNNs?cited by 187
  25. 2021Partial success in closing the gap between human and machine visioncited by 185
  26. 2023OpenOOD v1.5: Enhanced Benchmark for Out-of-Distribution Detectioncited by 178
  27. 2021Simple data balancing achieves competitive worst-group-accuracycited by 177
  28. 2021A Fine-Grained Analysis on Distribution Shiftcited by 171
  29. 2021Reading Race: AI Recognises Patient's Racial Identity In Medical Imagescited by 155
  30. 2022Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNscited by 155
  31. 2022FlexiViT: One Model for All Patch Sizescited by 151
  32. 2022On Feature Learning in the Presence of Spurious Correlationscited by 145
  33. 2023Getting aligned on representational alignmentcited by 142
  34. 2023Text-to-Image Diffusion Models are Zero-Shot Classifierscited by 140
  35. 2022Domain Generalization by Mutual-Information Regularization with Pre-trained Modelscited by 133
  36. 2021PixMix: Dreamlike Pictures Comprehensively Improve Safety Measurescited by 131
  37. 2021Advances in adversarial attacks and defenses in computer vision: A surveycited by 127
  38. 2022Toward Transparent AI: A Survey on Interpreting the Inner Structures of Deep Neural Networkscited by 126

50 most cited of 1,430 All 1,430 (JSON)