arxiv: 2604.25065 · v1 · submitted 2026-04-27 · 💻 cs.CV

Recognition: unknown

ShapeY: A Principled Framework for Measuring Shape Recognition Capacity via Nearest-Neighbor Matching

Jong Woo Nam , Amanda S. Rios , Bartlett W. Mel

Authors on Pith no claims yet

Pith reviewed 2026-05-08 04:09 UTC · model grok-4.3

classification 💻 cs.CV

keywords shape recognitionobject recognitiondeep networksbenchmarkingnearest-neighbor matchingviewpoint invariance3D shapeembedding space

0 comments

The pith

ShapeY shows that deep networks struggle to recognize objects by 3D shape across viewpoint and appearance changes.

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

The paper introduces ShapeY as a benchmarking framework to measure shape recognition capacity in object recognition systems using nearest-neighbor matching on rendered images of 3D objects. It evaluates 321 pre-trained networks and finds that even state-of-the-art models fail to consistently group different views of the same object by their 3D shape when viewpoints change or non-shape appearance alterations are introduced. This is important because human vision relies on shape cues for robust recognition, whereas deep networks often use texture and other shortcuts, leading to vulnerabilities. ShapeY provides various quantitative and qualitative metrics to assess the fine-grained structure of embedding spaces for shape understanding. The results highlight the need for better disentangled representations in artificial vision systems.

Core claim

ShapeY comprises 68,200 grayscale images of 200 3D objects rendered from multiple viewpoints, with optional non-shape appearance changes. Using a nearest-neighbor matching task, it probes whether an OR system's embedding space clusters object views by 3D shape similarity. Testing of 321 diverse pre-trained networks shows that state-of-the-art models struggle to generalize consistently across 3D viewpoint and appearance changes and are prone to infrequent but egregious matches of objects with completely different shapes. ShapeY thus establishes a principled framework for advancing toward human-like shape recognition capabilities.

What carries the argument

Nearest-neighbor matching distances in the embedding space to evaluate clustering of object views by 3D shape similarity across variations.

If this is right

Adoption of ShapeY would allow systematic comparison and improvement of shape invariance in new model architectures.
Well-performing models on ShapeY would exhibit better generalization in tasks requiring shape-based understanding despite visual changes.
The detailed readouts like viewpoint tuning curves can identify specific weaknesses in how models encode shape.
The framework emphasizes the development of encodings that are disentangled from non-shape cues like texture.
It provides a tool to track progress toward more robust and human-like object recognition systems.

Where Pith is reading between the lines

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

This finding implies that training data and objectives in current models may be insufficient for learning true 3D shape representations, suggesting a need for synthetic 3D data augmentation.
ShapeY could be used to evaluate emerging vision transformers or other architectures for their shape capabilities in ways standard datasets do not.
Poor shape recognition might underlie failures in applications like autonomous driving where viewpoint-invariant shape understanding is critical.
Extending ShapeY to include more complex scenes or real-world images could bridge the gap to practical deployment.

Load-bearing premise

That nearest-neighbor matching distances in a model's embedding space directly measure its capacity for shape-based recognition across viewpoint and appearance variation.

What would settle it

If a model achieves low error rates on ShapeY match tasks yet human observers rate the same objects as having dissimilar 3D shapes, or if the model succeeds on ShapeY but fails independent shape-based recognition tests outside the benchmark images.

Figures

Figures reproduced from arXiv: 2604.25065 by Amanda S. Rios, Bartlett W. Mel, Jong Woo Nam.

**Figure 1.** Figure 1: Overview of how ShapeY tests shape understanding in object view at source ↗

**Figure 2.** Figure 2: Complete set of 3D object models rendered using Blender at their “origin viewpoints”. Objects are grouped into 20 categories with 10 instances of view at source ↗

**Figure 4.** Figure 4: Example of a matching task. Given a reference image, we rank order view at source ↗

**Figure 6.** Figure 6: Examples of soft and hard contrast exclusions. In both soft and view at source ↗

**Figure 5.** Figure 5: Positive match candidates (PMCs) for view #8 (blue box) out of view at source ↗

**Figure 8.** Figure 8: Rank histograms showing the distribution of ranks for the top positive view at source ↗

**Figure 9.** Figure 9: Nearest-neighbor matching errors when a “contrast exclusion” is view at source ↗

**Figure 10.** Figure 10: Tuning curves showing how each of the image in the view at source ↗

**Figure 11.** Figure 11: Comparison of cosine similarity scores for top positive and negative match candidates. view at source ↗

**Figure 12.** Figure 12: Example image panel displaying six rows of: (left) the reference view at source ↗

**Figure 13.** Figure 13: shows the object and category-level matching accuracies at re = 2 in pr (higher is better). Contrary to our hypothesis, all the self-supervised methods perform worse than the supervised ResNet50. The best-performing self-supervised method is DINO, which still underperforms the supervised ResNet50 by nearly 20% in the object-level matching task. SimCLR, previously noted for being more shape-biased [46], al… view at source ↗

**Figure 15.** Figure 15: Fine-tuning results reporting nearest-neighbor matching accuracies at view at source ↗

**Figure 16.** Figure 16: Nearest-neighbor matching accuracies (re = 2 in pr) plotted against number of parameters of 321 different DN architectures pre-trained on ImageNet1k with an exception of a ViT trained with DINOv2, where authors curated a larger dataset themselves (142 million images) [37] (weights taken from timm [47]). We plot the scaling behaviors of three notable architectures: (left) ViT [48]; (middle) ConvNext [49] a… view at source ↗

**Figure 18.** Figure 18: ImageNet images (subset of 100,000), rank-ordered by correlation view at source ↗

**Figure 19.** Figure 19: (a) Nearest-neighbor matching errors for DINOv2 (re = 2 in pr). Examples show clear mismatches across categories that humans would not make (OCD errors). (b) Adversarial examples of ShapeY images altered to resemble a different object category in the embedding space. We also experimented with fine-tuning ResNet50, both in supervised and self-supervised manners, to see if including 3D view variations could… view at source ↗

read the original abstract

Object recognition (OR) in humans relies heavily on shape cues and the ability to recognize objects across varying 3D viewpoints. Unlike humans, deep networks often rely on non-shape cues such as texture and background, leading to vulnerabilities in generalization and robustness. To address this gap, we introduce ShapeY, a novel and principled benchmarking framework designed to evaluate shape-based recognition capability in OR systems. ShapeY comprises 68,200 grayscale images of 200 3D objects rendered from multiple viewpoints and optionally subjected to non-shape ``appearance'' changes. Using a nearest-neighbor matching task, ShapeY specifically probes the fine-grained structure of an OR system's embedding space by evaluating whether object views are clustered by 3D shape similarity across varying 3D viewpoints and other non-shape changes. ShapeY provides a suite of quantitative and qualitative performance readouts, including error rate graphs, viewpoint tuning curves, histograms of positive and negative matching scores, and grids showing ordered best matches, which together offer a comprehensive evaluation of an OR system's shape understanding capability. Testing of 321 pre-trained networks with diverse architectures reveals significant challenges in achieving robust shape-based recognition: even state-of-the-art models struggle to generalize consistently across 3D viewpoint and appearance changes, and are prone to infrequent but egregious matches of objects of obviously completely different shape. ShapeY establishes a principled framework for advancing artificial vision systems toward human-like shape recognition capabilities, emphasizing the importance of disentangled and invariant object encodings.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

ShapeY gives a new large-scale benchmark for shape invariance via NN matching on rendered objects, but the results are hard to read as proof of missing shape capacity without positive controls.

read the letter

The main thing here is a new benchmark called ShapeY built from 68,200 grayscale renders of 200 3D objects across multiple viewpoints, with optional appearance changes, scored by nearest-neighbor matching in embedding space to see if same-shape views cluster together. They apply it to 321 pre-trained networks and report that even strong models often fail to generalize by shape and occasionally make glaring mismatches. The work does well by scaling the test set and supplying several readouts, including error graphs, viewpoint curves, score histograms, and example match grids. That gives a practical way to inspect embedding structure beyond standard classification accuracy. The scale and the focus on 3D invariance with perturbations are the genuinely new pieces relative to earlier shape-bias tests. The soft spot is the missing validation for the metric itself. The claim that poor NN performance shows limited shape recognition assumes embedding distances mainly track 3D shape across changes, but the paper does not include a positive control, such as an explicit shape descriptor or mesh invariant that should succeed if the protocol is sound. Without that, or a negative control with scrambled shapes, the failures could reflect generic embedding properties or training biases rather than a specific absence of shape processing. The stress-test concern stands up on the evidence given. This is aimed at researchers working on texture bias, viewpoint robustness, or new vision benchmarks. A reader who wants a concrete dataset and protocol to measure these issues would get direct value from it. The empirical reach is large enough that it deserves a serious referee, who would probably ask for the controls and more on object selection and rendering details. I would send it to peer review rather than desk reject.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces ShapeY, a benchmarking framework with 68,200 grayscale images of 200 3D objects rendered from multiple viewpoints with optional non-shape appearance changes. It evaluates shape-based recognition capacity in 321 pre-trained networks via a nearest-neighbor matching task that probes whether embeddings cluster object views by 3D shape similarity, reporting that even state-of-the-art models struggle to generalize consistently across viewpoint and appearance variation and exhibit infrequent but egregious mismatches of dissimilar shapes.

Significance. If the nearest-neighbor procedure validly isolates shape capacity, the work supplies a scalable diagnostic tool with quantitative readouts (error graphs, viewpoint curves, score histograms) and qualitative match grids, plus large-scale testing across architectures; this could help steer development of more viewpoint- and appearance-invariant representations.

major comments (2)

[Evaluation Protocol] The central claim that poor NN performance demonstrates limited shape recognition rests on the unvalidated premise that embedding distances primarily encode 3D shape across viewpoints. No positive control is reported using an explicit shape descriptor (e.g., silhouette or mesh-invariant matching) that should succeed on the same stimuli, nor a negative control with shape-scrambled images; without these, failures could arise from generic embedding statistics rather than absence of shape understanding (see evaluation protocol and results sections).
[Methods] Quantitative tables and error-rate graphs are referenced but the manuscript provides insufficient detail on data-generation parameters (exact viewpoint sampling, appearance-change distributions, object selection) and on how the 321 networks were chosen or fine-tuned, making it impossible to assess whether the reported generalization failures are robust or protocol-dependent.

minor comments (2)

[Results] The abstract and results text use the term 'egregious matches' without a precise operational definition or frequency statistic; add a quantitative threshold or histogram bin for such cases.
Figure captions for the match grids should explicitly state the ordering criterion (e.g., cosine distance) and whether appearance changes were applied in the displayed examples.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed comments on our manuscript introducing the ShapeY framework. We address each major comment below with the strongest honest defense possible, proposing targeted revisions to improve validation and transparency without misrepresenting the original work.

read point-by-point responses

Referee: [Evaluation Protocol] The central claim that poor NN performance demonstrates limited shape recognition rests on the unvalidated premise that embedding distances primarily encode 3D shape across viewpoints. No positive control is reported using an explicit shape descriptor (e.g., silhouette or mesh-invariant matching) that should succeed on the same stimuli, nor a negative control with shape-scrambled images; without these, failures could arise from generic embedding statistics rather than absence of shape understanding (see evaluation protocol and results sections).

Authors: We agree that explicit positive and negative controls would provide stronger evidence that the nearest-neighbor failures specifically reflect limitations in shape encoding rather than generic embedding properties. The current protocol defines positives as different views of the same object and negatives as different objects, with qualitative match grids already showing egregious shape mismatches in some models. However, to directly address the concern, the revised manuscript will add a positive control using silhouette-based nearest-neighbor matching on the same grayscale stimuli (expected to succeed if shape cues dominate) and a negative control on shape-scrambled versions of the images (e.g., via randomized part textures or permutations) to show performance collapse when shape is disrupted. These will be reported in an expanded evaluation protocol section. revision: yes
Referee: [Methods] Quantitative tables and error-rate graphs are referenced but the manuscript provides insufficient detail on data-generation parameters (exact viewpoint sampling, appearance-change distributions, object selection) and on how the 321 networks were chosen or fine-tuned, making it impossible to assess whether the reported generalization failures are robust or protocol-dependent.

Authors: We acknowledge that greater detail on data generation and network selection is necessary for full reproducibility and robustness assessment. The original manuscript prioritized the framework description and aggregate results across 321 networks, but we agree this leaves gaps. In the revised version, the Methods section will be expanded to specify: exact viewpoint sampling (uniform azimuth 0-360° in 15° steps, elevation -30° to 30° in 10° steps), appearance-change distributions (parameters for texture randomization, lighting intensity, and background variation), object selection criteria (200 models chosen for shape diversity from ShapeNet and similar repositories), and network curation (321 models drawn from standard pre-trained repositories such as PyTorch Hub and TensorFlow Hub with no fine-tuning applied to preserve zero-shot evaluation). These additions will allow readers to evaluate protocol dependence. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical benchmark

full rationale

The paper introduces ShapeY as a new dataset of 68,200 rendered images and applies a nearest-neighbor matching procedure directly to the embedding spaces of 321 pre-trained networks. No mathematical derivations, predictions, or first-principles results are claimed. Central claims rest on empirical error rates, viewpoint curves, and match grids obtained from the benchmark itself, without any reduction to fitted parameters, self-definitional equations, or load-bearing self-citations. The framework is self-contained against external benchmarks and does not invoke uniqueness theorems or ansatzes from prior author work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claim rests on the assumption that nearest-neighbor geometry in embedding space faithfully reflects shape recognition capacity; no free parameters are fitted in the reported protocol, and no new physical or mathematical entities are postulated.

axioms (1)

domain assumption Nearest-neighbor matching in a model's embedding space measures shape similarity across viewpoint and appearance changes
This is the core premise of the evaluation task described in the abstract.

invented entities (1)

ShapeY benchmark framework no independent evidence
purpose: To quantify shape recognition capacity via nearest-neighbor matching
New dataset and protocol introduced by the paper; no independent evidence outside the work itself.

pith-pipeline@v0.9.0 · 5569 in / 1438 out tokens · 40370 ms · 2026-05-08T04:09:25.345911+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

55 extracted references · 9 canonical work pages · 4 internal anchors

[1]

The lateral occipital complex and its role in object recognition,

K. Grill-Spector, Z. Kourtzi, and N. Kanwisher, “The lateral occipital complex and its role in object recognition,”Vision Research, vol. 41, no. 10-11, pp. 1409–1422, May 2001

2001
[2]

Recognition-by-components: A theory of human image understanding

I. Biederman, “Recognition-by-components: A theory of human image understanding.”Psychological Review, vol. 94, no. 2, pp. 115–147, Apr. 1987

1987
[3]

Surface versus edge-based determinants of visual recognition,

I. Biederman and G. Ju, “Surface versus edge-based determinants of visual recognition,”Cognitive Psychology, vol. 20, no. 1, pp. 38–64, Jan. 1988

1988
[4]

Visual intelligence: How we create what we see

D. D. Hoffman, “Visual intelligence: How we create what we see.” Visual intelligence: How we create what we see., pp. xv, 294–xv, 294, 1998

1998
[5]

Representation of Perceived Object Shape by the Human Lateral Occipital Complex,

Z. Kourtzi, “Representation of Perceived Object Shape by the Human Lateral Occipital Complex,”Science, vol. 293, no. 5534, pp. 1506–1509, Aug. 2001

2001
[6]

ImageNet: A large-scale hierarchical image database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” in2009 IEEE Conference on Computer Vision and Pattern Recognition, Jun. 2009, pp. 248–255

2009
[7]

doi:10.48550/ARXIV.1811.12231

R. Geirhos, P. Rubisch, C. Michaelis, M. Bethge, F. A. Wich- mann, and W. Brendel, “ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness,” arXiv:1811.12231 [cs, q-bio, stat], Jan. 2019

work page arXiv 2019
[8]

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,” arXiv:1904.00760 [cs, stat], Mar. 2019

work page arXiv 1904
[9]

Deep convolutional networks do not classify based on global object shape,

N. Baker, H. Lu, G. Erlikhman, and P. J. Kellman, “Deep convolutional networks do not classify based on global object shape,”PLOS Compu- tational Biology, vol. 14, no. 12, p. e1006613, Dec. 2018

2018
[10]

Shape- Biased Learning by Thinking Inside the Box,

N. M ¨uller, C. G. M. Snoek, I. I. A. Groen, and H. S. Scholte, “Shape- Biased Learning by Thinking Inside the Box,” p. 2024.05.30.595526, Jun. 2024

2024
[11]

Are Vision Language Models Texture or Shape Biased and Can We Steer Them?

P. Gavrikov, J. Lukasik, S. Jung, R. Geirhos, B. Lamm, M. J. Mirza, M. Keuper, and J. Keuper, “Are Vision Language Models Texture or Shape Biased and Can We Steer Them?” Mar. 2024

2024
[12]

Emergence of Shape Bias in Convolutional Neural Networks through Activation Sparsity

T. Li, Y . Li, Z. Wen, and T. S. Lee, “Emergence of Shape Bias in Convolutional Neural Networks through Activation Sparsity.”
[13]

Assessing Shape Bias Property of Convolutional Neural Networks,

H. Hosseini, B. Xiao, M. Jaiswal, and R. Poovendran, “Assessing Shape Bias Property of Convolutional Neural Networks,” in2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), Jun. 2018, pp. 2004–20 048

2018
[14]

Shape or Texture: Understanding Discriminative Features in CNNs,

M. A. Islam, M. Kowal, P. Esser, S. Jia, B. Ommer, K. G. Derpanis, and N. Bruce, “Shape or Texture: Understanding Discriminative Features in CNNs,” Jan. 2021

2021
[15]

Teaching deep networks to see shape: Lessons from a simplified visual world,

C. Jarvers and H. Neumann, “Teaching deep networks to see shape: Lessons from a simplified visual world,” 2024

2024
[16]

Hidden in Plain Sight: Evaluating Abstract Shape Recognition in Vision-Language Models

A. Hemmat, A. Davies, T. A. Lamb, J. Yuan, P. Torr, A. Khakzar, and F. Pinto, “Hidden in Plain Sight: Evaluating Abstract Shape Recognition in Vision-Language Models.”
[17]

Does enhanced shape bias improve neural network robustness to common corruptions?

C. K. Mummadi, R. Subramaniam, R. Hutmacher, J. Vitay, V . Fischer, and J. H. Metzen, “Does enhanced shape bias improve neural network robustness to common corruptions?” Apr. 2021

2021
[18]

Can Biases in ImageNet Models Explain Generalization?

P. Gavrikov and J. Keuper, “Can Biases in ImageNet Models Explain Generalization?”
[19]

Columbia Object Image Library (COIL-100)

S. A. Nene, S. K. Nayar, and H. Murase, “Columbia Object Image Library (COIL-100).”
[20]

Learning methods for generic object recognition with invariance to pose and lighting,

Y . LeCun, Fu Jie Huang, and L. Bottou, “Learning methods for generic object recognition with invariance to pose and lighting,” inProceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., vol. 2. Washington, DC, USA: IEEE, 2004, pp. 97–104

2004
[21]

A large-scale hierarchical multi- view RGB-D object dataset,

K. Lai, L. Bo, X. Ren, and D. Fox, “A large-scale hierarchical multi- view RGB-D object dataset,” in2011 IEEE International Conference on Robotics and Automation, 2011, pp. 1817–1824

2011
[22]

BigBIRD: A large-scale 3D database of object instances,

A. Singh, J. Sha, K. S. Narayan, T. Achim, and P. Abbeel, “BigBIRD: A large-scale 3D database of object instances,” in2014 IEEE International Conference on Robotics and Automation (ICRA). Hong Kong, China: IEEE, May 2014, pp. 509–516

2014
[23]

iLab-20M: A Large-Scale Controlled Object Dataset to Investigate Deep Learning,

A. Borji, S. Izadi, and L. Itti, “iLab-20M: A Large-Scale Controlled Object Dataset to Investigate Deep Learning,” in2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, NV , USA: IEEE, Jun. 2016, pp. 2221–2230

2016
[24]

CORe50: A New Dataset and Benchmark for Continuous Object Recognition,

V . Lomonaco and D. Maltoni, “CORe50: A New Dataset and Benchmark for Continuous Object Recognition,” inProceedings of the 1st Annual Conference on Robot Learning. PMLR, Oct. 2017, pp. 17–26

2017
[25]

Objectron: A Large Scale Dataset of Object-Centric Videos in the Wild with Pose Annotations,

A. Ahmadyan, L. Zhang, J. Wei, A. Ablavatski, and M. Grundmann, “Objectron: A Large Scale Dataset of Object-Centric Videos in the Wild with Pose Annotations,” Dec. 2020

2020
[26]

Common Objects in 3D: Large-Scale Learning and Evaluation of Real-life 3D Category Reconstruction,

J. Reizenstein, R. Shapovalov, P. Henzler, L. Sbordone, P. Labatut, and D. Novotny, “Common Objects in 3D: Large-Scale Learning and Evaluation of Real-life 3D Category Reconstruction,” Sep. 2021

2021
[27]

PUG: Photorealistic and Semantically Controllable Synthetic Data for Representation Learning,

F. Bordes, S. Shekhar, M. Ibrahim, D. Bouchacourt, P. Vincent, and A. S. Morcos, “PUG: Photorealistic and Semantically Controllable Synthetic Data for Representation Learning,” Dec. 2023

2023
[28]

Benchmarking Neural Network Ro- bustness to Common Corruptions and Perturbations,

D. Hendrycks and T. Dietterich, “Benchmarking Neural Network Ro- bustness to Common Corruptions and Perturbations,” Mar. 2019

2019
[29]

Learning Robust Global Representations by Penalizing Local Predictive Power,

H. Wang, S. Ge, E. P. Xing, and Z. C. Lipton, “Learning Robust Global Representations by Penalizing Local Predictive Power,” Nov. 2019

2019
[30]

One-shot viewpoint invariance in matching novel objects,

I. Biederman and M. Bar, “One-shot viewpoint invariance in matching novel objects,”Vision Research, vol. 39, no. 17, pp. 2885–2899, Aug. 1999

1999
[31]

On the Semantics of a Glance at a Scene,

I. Biederman, “On the Semantics of a Glance at a Scene,” inPerceptual Organization. Routledge, 1981

1981
[32]

ShapeNet: An Information-Rich 3D Model Repository

A. X. Chang, T. Funkhouser, L. Guibas, P. Hanrahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su, J. Xiao, L. Yi, and F. Yu, “ShapeNet: An Information-Rich 3D Model Repository,” arXiv:1512.03012 [cs], Dec. 2015

work page internal anchor Pith review arXiv 2015
[33]

Untangling invariant object recognition,

J. J. DiCarlo and D. D. Cox, “Untangling invariant object recognition,” Trends in Cognitive Sciences, vol. 11, no. 8, pp. 333–341, Aug. 2007

2007
[34]

Deep Residual Learning for Image Recognition

K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual Learning for Image Recognition,”arXiv:1512.03385 [cs], Dec. 2015

work page internal anchor Pith review arXiv 2015
[35]

PyTorch Image Models,

R. Wightman, “PyTorch Image Models,” 2019

2019
[36]

A Simple Framework for Contrastive Learning of Visual Representations

T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A Simple Framework for Contrastive Learning of Visual Representations,”arXiv:2002.05709 [cs, stat], Jun. 2020

work page internal anchor Pith review arXiv 2002
[37]

DINOv2: Learning Robust Visual Fea- tures without Supervision,

M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski, “DINOv2: Learning Robust Visual Fea- tures without Su...

2024
[38]

Decoupled Contrastive Learning,

C.-H. Yeh, C.-Y . Hong, Y .-C. Hsu, T.-L. Liu, Y . Chen, and Y . LeCun, “Decoupled Contrastive Learning,” Jul. 2022

2022
[39]

Unsupervised Learning of Visual Features by Contrasting Cluster Assignments,

M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin, “Unsupervised Learning of Visual Features by Contrasting Cluster Assignments,” Jan. 2021

2021
[40]

VICReg: Variance-Invariance- Covariance Regularization for Self-Supervised Learning,

A. Bardes, J. Ponce, and Y . LeCun, “VICReg: Variance-Invariance- Covariance Regularization for Self-Supervised Learning,” Jan. 2022

2022
[41]

Momentum Contrast for Unsupervised Visual Representation Learning,

K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum Contrast for Unsupervised Visual Representation Learning,” Mar. 2020

2020
[42]

Bootstrap your own latent: A new approach to self-supervised Learning,

J.-B. Grill, F. Strub, F. Altch ´e, C. Tallec, P. H. Richemond, E. Buchatskaya, C. Doersch, B. A. Pires, Z. D. Guo, M. G. Azar, B. Piot, K. Kavukcuoglu, R. Munos, and M. Valko, “Bootstrap your own latent: A new approach to self-supervised Learning,” Sep. 2020

2020
[43]

Barlow Twins: Self-Supervised Learning via Redundancy Reduction,

J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny, “Barlow Twins: Self-Supervised Learning via Redundancy Reduction,” Jun. 2021. 16

2021
[44]

Emerging Properties in Self-Supervised Vision Transform- ers,

M. Caron, H. Touvron, I. Misra, H. J ´egou, J. Mairal, P. Bojanowski, and A. Joulin, “Emerging Properties in Self-Supervised Vision Transform- ers,” May 2021

2021
[45]

Understanding dimensional collapse in contrastive self-supervised learning.arXiv preprint arXiv:2110.09348, 2021

L. Jing, P. Vincent, Y . LeCun, and Y . Tian, “Understanding Dimensional Collapse in Contrastive Self-supervised Learning,”arXiv:2110.09348 [cs], Apr. 2022

work page arXiv 2022
[46]

On the surprising similarities between supervised and self-supervised models,

R. Geirhos, K. Narayanappa, B. Mitzkus, M. Bethge, F. A. Wichmann, and W. Brendel, “On the surprising similarities between supervised and self-supervised models,”arXiv:2010.08377 [cs, q-bio], Oct. 2020

work page arXiv 2010
[47]

ResNet strikes back: An improved training procedure in timm,

R. Wightman, H. Touvron, and H. J ´egou, “ResNet strikes back: An improved training procedure in timm,” Oct. 2021

2021
[48]

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

A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale,”arXiv:2010.11929 [cs], Jun. 2021

work page internal anchor Pith review arXiv 2010
[49]

A ConvNet for the 2020s,

Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, and S. Xie, “A ConvNet for the 2020s,” Mar. 2022

2022
[50]

ConvFormer: Combining CNN and Transformer for Medical Image Segmentation,

P. Gu, Y . Zhang, C. Wang, and D. Z. Chen, “ConvFormer: Combining CNN and Transformer for Medical Image Segmentation,” Nov. 2022

2022
[51]

2106.09681 , archivePrefix=

A. El-Nouby, H. Touvron, M. Caron, P. Bojanowski, M. Douze, A. Joulin, I. Laptev, N. Neverova, G. Synnaeve, J. Verbeek, and H. Jegou, “XCiT: Cross-Covariance Image Transformers,”arXiv:2106.09681 [cs], Jun. 2021

work page arXiv 2021
[52]

Going deeper with Image Transformers,

H. Touvron, M. Cord, A. Sablayrolles, G. Synnaeve, and H. J ´egou, “Going deeper with Image Transformers,” Apr. 2021

2021
[53]

Masked Autoencoders Are Scalable Vision Learners,

K. He, X. Chen, S. Xie, Y . Li, P. Doll ´ar, and R. Girshick, “Masked Autoencoders Are Scalable Vision Learners,” Dec. 2021

2021
[54]

Shortcut learning in deep neural networks,

R. Geirhos, J.-H. Jacobsen, C. Michaelis, R. Zemel, W. Brendel, M. Bethge, and F. A. Wichmann, “Shortcut learning in deep neural networks,”Nature Machine Intelligence, vol. 2, no. 11, pp. 665–673, Nov. 2020

2020
[55]

ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms,

W. Yang, B. Zhang, and O. Russakovsky, “ImageNet-OOD: Deciphering Modern Out-of-Distribution Detection Algorithms,” Mar. 2024. Jong Woo NamJong Woo Nam is a Ph.D. candi- date in the Neuroscience Graduate Program at the University of Southern California. He earned his B.Eng. in Applied Mathematics from Franklin W. Olin College of Engineering. His research ...

2024