Pith. sign in

REVIEW 2 major objections 4 minor 45 references

The Impact of the Single-Label Assumption in Image Recognition Benchmarking

T0 review · 2 major / 4 minor · reviewed 2026-08-11 · deepseek-v4-flash

Pith's one-line read The reported ImageNetV2 accuracy drop is substantially a single-label measurement artifact.

desk verdict The paper's qualitative claim is plausible and important, but the ASMA metric as defined cannot produce the reported numbers, so the headline result is currently unverifiable. read the letter →

arxiv 2412.18409 v2 pith:H5DXSS2V submitted 2024-12-24 cs.CV

classification cs.CV
keywords multi-labelevaluationImageNetV2accuracygapsingle-labelassumptionpredictioncapabilityvariabletop-kASMAPatchML
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

The paper tries to establish that a large part of the widely cited 11–14% top-1 accuracy drop on ImageNetV2 is a measurement artifact: single-label metrics penalize models for predicting correct secondary labels on images that genuinely contain several objects. Under the multi-label annotations used here, roughly 48% of ImageNetV2 images have more than one valid label, versus about 16% of ImageNetV1 images. When predictions are scored with a variable top-k that credits every valid label, and with the proposed ASMA metric, the average V1-to-V2 gap shrinks from 6–14% to 0–6%. The paper further introduces PatchML, a synthetic dataset that removes contextual co-occurrence cues, and reports that many single-label-trained models still rank several correct objects at the top. If this holds, standard ImageNet leaderboards have been understating the multi-label recognition ability of modern classifiers.

What carries the argument

Two evaluation instruments carry the argument. Variable top-k selects the top k predicted classes for each image, where k equals the number of valid labels in its ground truth, so a model is credited for each correct secondary label. ASMA divides images into subgroups by label count, computes per-label accuracy within each subgroup, and averages those subgroup accuracies, preventing the rare high-label-count images from dominating. PatchML is a synthetic diagnostic dataset built by cropping objects from ImageNet images and pasting them in random non-overlapping grid cells on black canvases, removing co-occurrence and context cues; it isolates whether multi-label recognition is genuine object-level behavior or a byproduct of scene statistics.

What would settle it

Take the ImageNetV2 images the paper counts as multi-label, have independent annotators list all visible objects, and recompute ASMA and the V1-to-V2 gap; if the gap returns above the 0–6% range or the multi-label prevalence estimate changes materially, the reported artifact is an artifact of the annotation protocol.

Watch

Extended reading notes

Core claim

The central discovery is that the ImageNetV2 degradation is substantially a measurement artifact: because ImageNetV2 has a higher incidence of multi-label images than ImageNetV1, top-1 accuracy penalizes models for predicting valid secondary labels. The authors quantify this by constructing variable top-k predictions, where k equals the number of ground-truth labels per image, and by introducing Aggregate Subgroup Model Accuracy (ASMA), which averages per-label accuracy across subgroups of images with the same label count. Across 315 ImageNet-pretrained models, the V1-to-V2 gap falls from 6–14% under top-1 to 0–11% under ReaL accuracy and to 0–6% under ASMA. PatchML, a synthetic dataset of randomly combined cropped objects on blank canvases, further shows that models trained with single-label supervision still rank multiple correct labels near the top, indicating latent multi-label capability beyond what standard metrics report.

Load-bearing premise

The load-bearing premise is that the list of valid objects used for ImageNetV2 is complete and correct for every image; if those annotations miss valid objects or include spurious ones, the measured gap reduction is an artifact of the annotation protocol.

Editorial extensions

If this is right

  • The common 11–14% ImageNetV2 drop should not be read as pure generalization failure; multi-label-aware metrics cut it to 0–6%.
  • Models with high top-1 accuracy are not necessarily the best at ranking multiple valid labels, so top-1 leaderboards can misrank models for multi-object scenes.
  • Training signals that go beyond single-label targets (dense token labels, masked image modeling, distillation from web-scale teachers) predict stronger multi-label capability; scale alone is not enough.
  • Evaluation protocols for ImageNet should adopt multi-label-aware metrics alongside top-1 to avoid underestimating DNN capabilities.

Reading between the lines

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

  • A testable extension: apply the same variable top-k and ASMA protocol to other single-label benchmarks that likely contain multi-label images, such as Places or iNaturalist, to see whether their apparent domain gaps also narrow.
  • Beyond the paper's evaluation-only scope, the PatchML construction could be turned into a training curriculum; if models fine-tuned on multi-object composites improve ASMA on natural images, that would confirm the latent-capability story causally.
  • If the gap reduction is real, earlier conclusions that ImageNet classifiers fail under distribution shift may need to be re-stated as conclusions about single-label evaluation rather than about the models themselves.
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

2 major / 4 minor

Summary. The paper argues that the commonly reported 11–14% top-1 accuracy drop on ImageNetV2 is largely an artifact of the single-label evaluation protocol. It proposes variable top-k evaluation with k equal to the number of valid labels per image, introduces Aggregate Subgroup Model Accuracy (ASMA) as a subgroup-averaged multi-label accuracy, and presents PatchML, a synthetic dataset of composed object patches. Across 315 ImageNet-pretrained models, the authors report that the V1–V2 gap narrows to 0–6% under ASMA and that many models rank multiple correct labels highly. The paper concludes that single-label metrics understate DNN multi-label capability.

Significance. If the quantitative claims could be verified, the work would be valuable: it addresses a widely discussed anomaly in benchmarking, uses a broad model pool (315 TIMM models), proposes a synthetic diagnostic that removes co-occurrence cues, and ships a code repository (claimed). The reframing of the ImageNetV2 gap as partly evaluation-induced is consistent with prior work on ReaL labels and would have implications for benchmark design. However, the central quantitative claim currently rests on an ASMA definition that appears inconsistent with the reported numbers, and on self-authored ImageNetV2 annotations that lack independent validation; both need to be resolved before the conclusions can be accepted.

major comments (2)
  1. [Section 3.3, Eq. (Ag) and Figure 4c/Figure 5b] The ASMA definition is internally inconsistent with the reported numbers. The label-wise accuracy formula (1/C) * sum_c I(y_gt_g,i,c = y_hat_g,i,c) with C=1000 and y_hat in {0,1}^C is a Hamming-style accuracy, but no rule is given for constructing the binary prediction vector y_hat from the variable top-k softmax outputs of Section 3.1. If y_hat is the indicator of the top-k predicted classes and k equals the number of true labels g, then the per-image score is 1 - (false positives)/1000, which is at least ~0.99 for any image with all true labels in the top-k and typically not much lower otherwise. This cannot produce the subgroup accuracies near 0.2–0.7 in Figure 5b or the 0–6% ASMA gap in Figure 4c. The authors must correct the formula (e.g., to example-based accuracy |Y ∩ Yhat| / |Y ∪ Yhat|, which they claim to extend) or specify the binarization and recompute all ASMA-based figures. Because the abstract and Section 1.1 base the central 'gap shrinks to 0–6%' claim on ASMA, this inconsistency is load-bearing.
  2. [Section 4.2.1, Table 1, Figure 4] The quantitative gap reduction on ImageNetV2 depends entirely on the multi-label annotations from Anzaku et al. [14], which are self-authored and not machine-checked, reproduced, or independently validated in this paper. The paper provides no inter-annotator agreement, error analysis, or third-party comparison for these labels. Section 5 acknowledges that the annotations 'may still miss plausible object categories,' but the main conclusion that the V2 gap is an evaluation artifact assumes the labels are complete and correct. I ask for a sensitivity analysis or an independent validation subset (e.g., labels from a different annotation protocol, or agreement rates on a sample) before the gap-reduction claim is accepted.
minor comments (4)
  1. [Figure 4 caption] The caption says 'all plotted against label count' but the x-axis is the index of pre-trained models; please correct the caption to match the actual axes.
  2. [Section 3.3] The ASMA sum runs over g = 0 to G-1, but no images have zero labels in Table 1; please specify how empty subgroups are handled or start the sum at g = 1.
  3. [Section 3.2, Algorithm 1] It is unclear whether the patch pool S is reset for each k value and for each of the five seeds; please clarify, as this affects the independence of the PatchML variants summarized in Table 2.
  4. [Abstract and Section 3.3] The metric is called 'Aggregate Subgroup Model Accuracy' in the abstract and 'Average Subgroup Multi-Label Accuracy' in Section 3.3; please use a single consistent name throughout.

Circularity Check

1 steps flagged · score 4.0 of 10

The central ImageNetV2 gap-reduction result is measured against the authors' own unvalidated multi-label annotations (Anzaku et al. [14]); the ASMA formula in Section 3.3 is internally inconsistent with Figure 4c, a verifiability problem that is not itself a circular reduction.

  1. self citation load bearing [Section 4.1 (Test Dataset Description) and Section 4.2.1 (Prevalence of Multi-label Images), Table 1 and Figure 3]
    "For ImageNetV2, we used the refined labels generated by Anzaku et al. [14], which account for multi-label images. ... To do this, we utilize the re-assessed labels from ReaL [5] for ImageNetV1 and the refined labels introduced by Anzaku et al. [14] for ImageNetV2."

    The paper's headline finding—that ImageNetV2 is approximately 48% multi-label versus 16% for ImageNetV1, and that the top-1 accuracy drop shrinks to 0–6% under ASMA—is computed against a ground-truth label set created by the same authors (Anzaku et al. [14]). No independent human agreement, machine check, or external validation of these labels is provided, and Section 5 concedes they 'may still miss plausible object categories.' The central quantitative claim therefore inherits its value from the authors' own annotation protocol rather than from an independently verified measurement.

full rationale

The paper is not globally circular: the ReaL labels for ImageNetV1 are external, the PatchML diagnostic is a freshly constructed controlled dataset with labels derived from object-detection annotations, and the ranking analysis of 315 TIMM models is self-contained. Those parts provide independent content. The main circularity concern is the load-bearing reliance on the authors' own ImageNetV2 annotations from Anzaku et al. [14] for both the multi-label prevalence claim and the headline 0–6% ASMA gap. Because these labels are neither machine-checked nor independently validated in the paper, the central V2-specific result is not externally anchored. Separately, the ASMA definition in Section 3.3 is inconsistent with Figure 4c: the stated Hamming-style label-wise accuracy over C=1000 classes, combined with the variable top-k prediction mechanism of Section 3.1, lacks a rule converting top-k indices to the binary vector y_hat, and under the natural conversion every model would score near (C-k)/C, making the plotted 0–0.9-scale values and 0–6% gap unreproducible. This is a definitional/verifiability flaw, not a circular reduction to the paper's own inputs, so it is flagged here rather than counted as a circular step. On balance the self-cited annotation input is substantial and load-bearing but not the whole derivation, yielding a score of 4.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on two label sets (ReaL for ImageNetV1 and the authors' own ImageNetV2 annotations), on the top-k ranking assumption, and on PatchML label completeness. No fitted numerical parameters underlie the metrics themselves, though the PatchML composition schedule is chosen by hand. The self-authored ImageNetV2 label set is the main circularity burden.

free parameters (2)
  • PatchML object counts K = {2, 3, 4, 6, 9}
    Chosen patch counts per composite image; affects test difficulty and the subgroup accuracy values in Figure 6, but not fitted to model outputs.
  • PatchML patch sizes P = {256, 256, 256, 170, 128}
    Chosen to fit the 512x512 canvas; larger k uses smaller patches to keep all objects visible.
assumptions (4)
  • domain assumption The ReaL multi-label annotations for ImageNetV1 are complete and correct for the 46,837 images they cover.
    Used as ground truth for all multi-label metrics on ImageNetV1 (Section 4.1). Missing or erroneous labels would change the measured multi-label capability.
  • domain assumption The ImageNetV2 multi-label annotations from Anzaku et al. [14] are complete and correct for the 9,858 annotated images.
    Used as ground truth for ImageNetV2 (Section 4.1, Table 1). These annotations are author-produced and not independently validated in this paper, making them the most fragile premise.
  • domain assumption A model that recognizes all valid objects in an image will place those objects among its top-k predictions, where k is the number of valid labels.
    Explicitly stated as an assumption of variable top-k evaluation in the Limitations section; if false, the ASMA and PatchML analyses would misestimate multi-label recognition.
  • domain assumption The PatchML labels, obtained by taking the union of source-patch labels, are the complete set of recognizable objects in each composite image.
    Used to score PatchML subgroup accuracy; black backgrounds and cropped patches are assumed to introduce no additional valid labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Impact of the Single-Label Assumption in Image Recognition Benchmarking." pith.science (2026). https://pith.science/paper/H5DXSS2V

@misc{pith2026241218409,
  author       = {Pith},
  title        = {Pith review of: The Impact of the Single-Label Assumption in Image Recognition Benchmarking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H5DXSS2V}},
  note         = {Machine review of arXiv:2412.18409}
}
abstract

Deep neural networks (DNNs) are typically evaluated under the assumption that each image has a single correct label. However, many images in benchmarks like ImageNet contain multiple valid labels, creating a mismatch between evaluation protocols and the actual complexity of visual data. This mismatch can penalize DNNs for predicting correct but unannotated labels, which may partly explain reported accuracy drops, such as the widely cited 11 to 14 percent top-1 accuracy decline on ImageNetV2, a replication test set for ImageNet. This raises the question: do such drops reflect genuine generalization failures or artifacts of restrictive evaluation metrics? We rigorously assess the impact of multi-label characteristics on reported accuracy gaps. To evaluate the multi-label prediction capability (MLPC) of single-label-trained models, we introduce a variable top-$k$ evaluation, where $k$ matches the number of valid labels per image. Applied to 315 ImageNet-trained models, our analyses demonstrate that conventional top-1 accuracy disproportionately penalizes valid but secondary predictions. We also propose Aggregate Subgroup Model Accuracy (ASMA) to better capture multi-label performance across model subgroups. Our results reveal wide variability in MLPC, with some models consistently ranking multiple correct labels higher. Under this evaluation, the perceived gap between ImageNet and ImageNetV2 narrows substantially. To further isolate multi-label recognition performance from contextual cues, we introduce PatchML, a synthetic dataset containing systematically combined object patches. PatchML demonstrates that many models trained with single-label supervision nonetheless recognize multiple objects. Altogether, these findings highlight limitations in single-label evaluation and reveal that modern DNNs have stronger multi-label capabilities than standard metrics suggest.

Figures

Figures reproduced from arXiv: 2412.18409 by the authors.

Figure 1
Figure 1. Examples from ImageNetV2 showing the top-5 predictions of a pre-trained DNN model. The [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. An illustration of the PatchML dataset creation process, organized into two main stages: (1) [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Distribution of images by number of ground-truth labels in ImageNetV1 and ImageNetV2, with [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Accuracies obtained by assessing 315 pre-trained DNNs on [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Effectiveness distribution of 315 DNNs pre-trained on ImageNet, with effectiveness obtained on ImageNetV1 and ImageNetV2 (for datapoints with 1-5 labels): (a) ReaL accuracy and (b) Subgroup accuracy versus Label Count. Each dot represents the accuracy of a model for im…
Figure 6
Figure 6. Figure 6: Subgroup accuracy boxplot for 315 pre-trained DNNs evaluated on PatchML. Overlaid are line [PITH_FULL_IMAGE:figures/full_fig_p020_6.png]
Figure 7
Figure 7. Figure 7: Examples of PatchML predictions for sample images with nine ground truth labels. Red bars [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 31 canonical work pages

  1. [14]

    E. T. Anzaku, H. Hong, J.-W. Park, W. Yang, K. Kim, J. Won, D. V . K. Herath, A. Van Messem, W. De Neve, Leveraging Human-Machine Interactions for Com- puter Vision Dataset Quality Enhancement, in: B. J. Choi, D. Singh, U. S. Tiwary, W.-Y . Chung (Eds.), Intelligent Human Computer Interaction, 2024, pp. 295–309. doi:10.1007/978-3-031-53827-8_27

  2. [1]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, Imagenet: A Large- scale Hierarchical Image Database, in: IEEE conference on computer vision and pattern recognition, 2009, pp. 248–255

  3. [2]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, L. Fei-Fei, ImageNet Large Scale Visual Recognition Challenge, International Journal of Computer Vision 115 (2015) 211–252. doi:10.1007/s11263-015-0816-y

  4. [3]

    Barbu, D

    A. Barbu, D. Mayo, J. Alverio, W. Luo, C. Wang, D. Gutfreund, J. Tenenbaum, B. Katz, ObjectNet: A large-scale bias-controlled dataset for pushing the limits of object recognition models, in: Advances in Neural Information Processing Systems, volume 32, 2019. 28

  5. [4]

    Recht, R

    B. Recht, R. Roelofs, L. Schmidt, V . Shankar, Do ImageNet Classifiers Gener- alize to ImageNet?, in: K. Chaudhuri, R. Salakhutdinov (Eds.), Proceedings of the 36th International Conference on Machine Learning, volume 97, 2019, pp. 5389–5400

  6. [5]

    Beyer, O

    L. Beyer, O. J. H ´enaff, A. Kolesnikov, X. Zhai, A. v. d. Oord, Are we done with ImageNet?, 2020. URL: http://arxiv.org/abs/2006.07159

  7. [6]

    Shankar, R

    V . Shankar, R. Roelofs, H. Mania, A. Fang, B. Recht, L. Schmidt, Evaluating Ma- chine Accuracy on ImageNet, in: International Conference on Machine Learning, volume 37, 2020, pp. 8634–8644

  8. [7]

    Tsipras, S

    D. Tsipras, S. Santurkar, L. Engstrom, A. Ilyas, A. Madry, From ImageNet to Image Classification: Contextualizing Progress on Benchmarks, in: International Conference on Machine Learning, 2020, pp. 9625–9635

Show all 45 references
  1. [8]

    Vasudevan, B

    V . Vasudevan, B. Caine, R. Gontijo-Lopes, S. Fridovich-Keil, R. Roelofs, When does dough become a bagel? Analyzing the remaining mistakes on ImageNet, in: Conference on Neural Information Processing Systems, 2022

  2. [9]

    Kisel, I

    N. Kisel, I. V olkov, K. Hanzelkova, K. Janouskova, J. Matas, Flaws of Im- ageNet, Computer Vision’s Favourite Dataset, 2024. doi: 10.48550/arXiv. 2412.00076

  3. [10]

    C. G. Northcutt, A. Athalye, J. Mueller, Pervasive Label Errors in Test Sets Destabilize Machine Learning Benchmarks, in: Thirty-fifth Conference on Neu- ral Information Processing Systems Datasets and Benchmarks Track, 2021. URL: https://openreview.net/forum?id=XccDXrDNLek

  4. [11]

    A. S. Luccioni, D. Rolnick, Bugs in the data: how ImageNet misrepresents bio- diversity, in: Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence and Thirteenth Symposium ...

  5. [12]

    B. Y . Idrissi, D. Bouchacourt, R. Balestriero, I. Evtimov, C. Hazirbas, N. Ballas, P. Vincent, M. Drozdzal, D. Lopez-Paz, M. Ibrahim, ImageNet-X: Understanding Model Mistakes with Factor of Variation Annotations, 2022

  6. [13]

    Peychev, M

    M. Peychev, M. N. M ¨uller, M. Fischer, M. Vechev, Automated Classification of Model Errors on ImageNet, Thirty-seventh Conference on Neural Information Processing Systems (2023)

  7. [15]

    Stock, M

    P. Stock, M. Cisse, ConvNets and ImageNet Beyond Accuracy: Understanding Mistakes and Uncovering Biases, in: The European Conference on Computer Vision, 2018, pp. 504–519. doi:10.1007/978-3-030-01231-1_31

  8. [16]

    Engstrom, A

    L. Engstrom, A. Ilyas, S. Santurkar, D. Tsipras, J. Steinhardt, A. Madry, Identi- fying Statistical Bias in Dataset Replication, in: Proceedings of the 37th Interna- tional Conference on Machine Learning, volume 119, 2020, pp. 2922–2932

  9. [17]

    E. T. Anzaku, H. Wang, A. Babalola, A. Van Messem, W. De Neve, Re- assessing accuracy degradation: a framework for understanding DNN be- havior on similar-but-non-identical test datasets, Machine Learning 114 (2025). URL: https://doi.org/10.1007/s10994-024-06693-x . doi: 10. 1...

  10. [18]

    M. R. Taesiri, G. Nguyen, S. Habchi, C.-P. Bezemer, A. Nguyen, ImageNet-Hard: The Hardest Images Remaining from a Study of the Power of Zoom and Spatial Biases in Image Classification, in: Conference on Neural Information Processing Systems, 2023

  11. [19]

    E. Cole, O. M. Aodha, T. Lorieul, P. Perona, D. Morris, N. Jojic, Multi-Label Learning from Single Positive Labels, in: IEEE /CVF Conference on Computer 30 Vision and Pattern Recognition, 2021, pp. 933–942. doi: 10.1109/CVPR46437. 2021.00099

  12. [20]

    Verelst, P

    T. Verelst, P. K. Rubenstein, M. Eichner, T. Tuytelaars, M. Berman, Spatial Consistency Loss for Training Multi-Label Classifiers from Single-Label Anno- tations, in: IEEE /CVF Winter Conference on Applications of Computer Vision, 2023, pp. 3868–3878. doi:10.1109/WACV56688.2023.00387

  13. [21]

    S. Yun, S. J. Oh, B. Heo, D. Han, J. Choe, S. Chun, Re-labeling ImageNet: from Single to Multi-Labels, from Global to Localized Labels, in: The IEEE /CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 2340–2350. doi:10.1109/CVPR46437.2021.00237

  14. [22]

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, Y . Yoo, CutMix: Regularization Strategy to Train Strong Classifiers with Localizable Features, in: International Conference on Computer Vision (ICCV), 2019

  15. [23]

    Takahashi, T

    R. Takahashi, T. Matsubara, K. Uehara, RICAP: Random Image Cropping and Patching Data Augmentation for Deep CNNs, in: Proceedings of The 10th Asian Conference on Machine Learning, volume 95, 2018, pp. 786–798

  16. [24]

    L. Deng, Y . Liu, Y . Shi, W. Zhang, C. Yang, H. Liu, Deep neural networks for inferring binding sites of RNA-binding proteins by using distributed repre- sentations of RNA primary sequence and secondary structure, BMC Genomics 21 (2020) 866. URL: https://doi.org/10.1186/s1286...

  17. [25]

    Zhang, Z.-H

    M.-L. Zhang, Z.-H. Zhou, A Review on Multi-Label Learning Algorithms, IEEE Transactions on Knowledge and Data Engineering 26 (2014) 1819–1837. doi:10. 1109/TKDE.2013.39

  18. [26]

    Wightman, PyTorch Image Models, 2019

    R. Wightman, PyTorch Image Models, 2019. URL: https://github.com/ rwightman/pytorch-image-models. doi:10.5281/zenodo.4414861, publi- cation Title: GitHub repository. 31

  19. [27]

    K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image Recognition, in: IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778. doi:10.1109/CVPR.2016.90

  20. [28]

    A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand, M. An- dreetto, H. Adam, MobileNets: Efficient Convolutional Neural Networks for Mo- bile Vision Applications, 2017. URL: http://arxiv.org/abs/1704.04861. doi:10.48550/arXiv.1704.04861

  21. [29]

    M. Tan, Q. V . Le, EfficientNet: Rethinking Model Scaling for Convolutional Neu- ral Networks, in: Proceedings of the 36th International Conference on Machine Learning, volume 97, 2019, pp. 6105–6114

  22. [30]

    S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, S. Xie, Con- vNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 16133–16142. doi:10.1109/CVPR52729.2023.01548

  23. [31]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, N. Houlsby, An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale, in: Ninth International Conference on Learnin...

  24. [32]

    Touvron, M

    H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, H. J ´egou, Training Data-efficient Image Transformers & Distillation Through Attention, in: Pro- ceedings of the 38th International Conference on Machine Learning, volume 139, 2021, pp. 10347–10357

  25. [33]

    L. Dong, S. Piao, F. Wei, BEiT: BERT Pre-Training of Image Transformers, in: The Tenth International Conference on Learning Representations, 2022

  26. [35]

    Touvron, M

    H. Touvron, M. Cord, A. Sablayrolles, G. Synnaeve, H. Jegou, Going deeper with Image Transformers, in: IEEE /CVF International Conference on Computer Vision (ICCV), IEEE, 2021, pp. 32–42. doi:10.1109/ICCV48922.2021.00010

  27. [36]

    L. Yuan, Q. Hou, Z. Jiang, J. Feng, S. Yan, VOLO: Vision Outlooker for Visual Recognition, IEEE Transactions on Pattern Analysis and Machine Intelligence 45 (2023) 6575–6586. URL: https://ieeexplore.ieee.org/ document/9888055. doi:10.1109/TPAMI.2022.3206108

  28. [37]

    Q. Sun, Y . Fang, L. Wu, X. Wang, Y . Cao, EV A-CLIP: Improved Training Tech- niques for CLIP at Scale, 2023. doi:10.48550/arXiv.2303.15389

  29. [38]

    Mahajan, K

    D. Mahajan, K. He, M. Paluri, Y . Li, A. Bharambe, L. Van Der Maaten, Explor- ing the Limits of Weakly Supervised Pretraining, in: European Conference on Computer Vision, volume 11206, 2018, pp. 185–201

  30. [39]

    Singh, L

    M. Singh, L. Gustafson, A. Adcock, V . De Freitas Reis, B. Gedik, R. P. Kosaraju, D. Mahajan, R. Girshick, P. Dollar, L. Van Der Maaten, Revisiting Weakly Supervised Pre-Training of Visual Perception Models, in: IEEE /CVF Conference on Computer Vision and Pattern Recognition, ...

  31. [40]

    Jiang, Q

    Z.-H. Jiang, Q. Hou, L. Yuan, D. Zhou, Y . Shi, X. Jin, A. Wang, J. Feng, All Tokens Matter: Token Labeling for Training Better Vision Transformers, in: Ad- vances in Neural Information Processing Systems, volume 34, 2021, pp. 18590– 18602

  32. [41]

    Z. Liu, H. Mao, C.-Y . Wu, C. Feichtenhofer, T. Darrell, S. Xie, A ConvNet for the 2020s, in: IEEE Conference on Computer Vision and Pattern Recognition, 2022, pp. 11976–11986

  33. [42]

    Z. Peng, L. Dong, H. Bao, Q. Ye, F. Wei, BEiT v2: Masked Image Model- ing with Vector-Quantized Visual Tokenizers, 2022. URL:http://arxiv.org/ abs/2208.06366. doi:10.48550/arXiv.2208.06366. 33

  34. [43]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, I. Sutskever, Learning Transferable Visual Models From Natural Language Supervision, in: Proceedings of the 38th International Conference on Machine Learni...

  35. [44]

    Radosavovic, R

    I. Radosavovic, R. P. Kosaraju, R. Girshick, K. He, P. Dollar, Designing Net- work Design Spaces, in: IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 10425–10433. doi:10.1109/CVPR42600.2020. 01044

  36. [45]

    S. S. S. Kruthiventi, K. Ayush, R. V . Babu, DeepFix: A Fully Convolutional Neu- ral Network for predicting Human Eye Fixations, 2015. URL: http://arxiv. org/abs/1510.02927. doi:10.48550/arXiv.1510.02927

  37. [46]

    Jetley, N

    S. Jetley, N. Murray, E. Vig, End-to-End Saliency Mapping via Probability Distri- bution Prediction, in: IEEE Conference on Computer Vision and Pattern Recog- nition, 2016, pp. 5753–5761. doi:10.1109/CVPR.2016.620. 34

Pith tools

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