Pith. sign in

REVIEW 4 major objections 6 minor 33 references

Beyond Linear Bottlenecks: Spline-Based Knowledge Distillation for Culturally Diverse Art Style Classification

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Replacing MLP projection heads with spline-based KAN heads raises Top-1 art-style classification accuracy on the WikiArt and Pandora18k benchmarks.

desk verdict Plausible KAN-head swap in a dual-teacher SSL pipeline, but the reported gains are within seed noise and the paper contradicts itself on which architecture it actually evaluates. read the letter →

arxiv 2507.23436 v1 pith:BPGCXKO4 submitted 2025-07-31 cs.CV

Abdellah Zakaria Sellam , Salah Eddine Bekhouche , Cosimo Distante , Abdelmalik Taleb-Ahmed This is my paper · ORCID
classification cs.CV
keywords artstyleclassificationknowledgedistillationdual-teacherKolmogorov-ArnoldNetworksself-supervisedlearningGrammatrixalignmentsplineactivationsfine-grained
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

Art style classification suffers from a shortage of expert labels and from stylistic cues that interact nonlinearly. This paper argues that existing dual-teacher self-supervised frameworks leave accuracy on the table by using fixed-activation MLP projection heads, and proposes replacing those heads with Kolmogorov–Arnold Networks (KANs), whose spline-based learnable activations can represent high-order feature interactions. On WikiArt and Pandora18k, the authors report Top-1 gains from KAN heads on most backbone–dataset pairs (about 0.2 to 1.0 percentage points, with EfficientNet-B0 on WikiArt essentially flat), plus higher linear probe accuracy than MLP projections. If the claim holds, swapping projection heads is a low-friction way to extract more style signal from unlabeled art data without changing the encoders or the loss family.

What carries the argument

The central object is the Kolmogorov–Arnold Network (KAN) projection head, a neural layer grounded in the Kolmogorov–Arnold representation theorem in which each edge is a learnable univariate cubic B-spline with trainable control points and adaptive knot spacing instead of a fixed activation. In this pipeline, every branch (student, momentum teacher, style teacher) projects encoder features through KAN heads; the resulting embeddings feed cosine-similarity relation alignment and KL-divergence losses, while Gram matrices computed from encoder features enforce texture and style agreement. The KAN head is regularized by spline-parameter sparsity, second-derivative smoothness, and segment dropout. This object carries the argument because it is the only component changed between the 'Base' and 'KAN' rows in the comparison tables.

What would settle it

Retrain every Base and KAN configuration in Tables 1 and 2 with five different random seeds under identical hyperparameters, and plot mean Top-1 with confidence intervals. The central claim stands only if the KAN intervals clear the Base intervals on most rows; if the intervals overlap, the reported 0.03 to 1.26 point deltas are indistinguishable from run-to-run noise.

Watch

Extended reading notes

Core claim

The central discovery is that the projection head, not just the encoder, limits how separable a self-supervised art-style representation can be. The authors take the dual-teacher distillation pipeline—one momentum teacher, one style teacher, and a student—and replace the MLP projection and prediction heads with KAN heads in which each edge is a learnable cubic B-spline function with adaptive knot spacing. In their tables this yields Pandora18k Top-1 gains of 0.92 points for EfficientNet-B0, 1.03 for ConvNeXt-Base, and 0.39 for ViT-Base; on WikiArt, ConvNeXt-Base gains 0.87, ViT-Base gains 0.23, and EfficientNet-B0 is flat at -0.03. A placement study on ConvNeXt-Base/Pandora18k finds the largest single change (+1.26) when only the style-teacher head is KAN, while KAN in the momentum-teacher head alone adds little. The paper interprets this as spline activations disentangling overlapping style manifolds more cleanly than fixed linear projections.

Load-bearing premise

The load-bearing premise is that Top-1 differences of roughly 0.03 to 1.26 percentage points, each measured on a single training run with no error bars or significance test, reflect genuine improvements rather than random seed or hyperparameter noise.

Editorial extensions

If this is right

  • KAN projection heads are a drop-in replacement: the EMA teacher updates, Gram-matrix style loss, and relation alignment loss remain unchanged from the base dual-teacher pipeline.
  • Where the KAN head is placed matters: putting it in the style-teacher branch alone gives the largest Top-1 gain, while the momentum-teacher branch alone gives almost nothing.
  • KAN heads improve linear probe accuracy, which means the gains are in the learned embedding geometry rather than in the final classifier.
  • The benefit is not uniform across backbones: on WikiArt, ConvNeXt-Base and ViT-Base improve while EfficientNet-B0's Top-1 is flat, so the effect depends on the encoder and dataset.
  • Remaining errors concentrate in styles that require cultural context, such as Ukiyo-e and Socialist Realism, indicating texture-level improvements do not resolve all semantic ambiguity.

Reading between the lines

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

  • Because each table row reports one training run, a multi-seed retest is the natural next step: the mechanism claim is strongest where the KAN–base gap is about one percentage point, and weakest for the 0.03 to 0.4 point deltas.
  • Varying the KAN grid size and spline order while holding everything else fixed would test whether accuracy tracks spline capacity; if it does not, the gains might come from regularization or added parameters rather than nonlinear expressiveness.
  • The same head swap could plausibly transfer to other weak-label, fine-grained visual domains where texture and composition, not object identity, define the classes; this is untested in the paper.
  • The confusion-matrix evidence suggests culturally loaded styles remain hard, so combining KAN features with artist, period, or geographic metadata is a promising but untested extension.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes replacing the MLP projection and prediction heads in a dual-teacher self-supervised knowledge-distillation framework for art style classification with spline-based Kolmogorov-Arnold Network (KAN) heads. It reports Top-1 accuracy comparisons against an MLP baseline on WikiArt and Pandora18k across EfficientNet-B0, ConvNeXt-Base, and ViT-Base, and includes an ablation over KAN placement in the projection heads. The central claim is that KAN heads improve Top-1 accuracy over the base dual-teacher architecture.

Significance. The motivation is reasonable: KANs can model nonlinear feature interactions that fixed-activation MLPs may miss, and the dual-teacher setup for art-style classification is a sensible testbed. If the reported gains were reproducible and statistically significant, the paper would be a useful incremental contribution to self-supervised representation learning for art. However, the manuscript provides no code link, no error bars, no significance tests, and no parameter-free derivation; the contribution is entirely empirical. In its current form, the evidence does not establish the central claim, although the underlying idea remains worth investigating.

major comments (4)
  1. [§3.1, §4, Table 2] The evaluated architecture is not well-defined. §3.1 states that 'all branches use a KAN as their projection function' and Figure 1 labels all three projection heads as KAN, but §4 says the framework trains a student using one teacher with an MLP head and another with a KAN head. Table 2 then lists 'Style Teacher KAN' at 66.49 Top-1, higher than 'All Heads KAN (Ours)' at 66.26, even though the text calls the all-heads variant 'most significant.' Because Table 1's 'KAN' rows are not tied to a specific head configuration, the headline comparison is ambiguous and the proposed method's claimed advantage is not established.
  2. [§4.3, Tables 1 and 2] Every reported Top-1 value comes from a single run, with no standard deviation, confidence interval, or significance test. The differences between KAN and base range from -0.03 to +1.26 percentage points; for example, EfficientNet-B0 on WikiArt drops from 50.09 to 50.06. Since run-to-run variation in self-supervised training is typically non-negligible at these margins, the abstract's unconditional claim that the approach 'outperforms the base dual teacher architecture' is not supported by the data as presented.
  3. [§4.2, Eq. (5)] The KAN regularization weights λ_L1 and λ_smooth and the segment deactivation rate are never specified, and the KAN width/depth are not reported. Without these values, the comparison with the MLP baseline cannot be reproduced, and it is impossible to verify that the two architectures are matched in capacity and training budget. The text promises a public repository but gives no link; the reproducibility of the empirical claim is therefore incomplete.
  4. [§5.1, Table 2] The text says the 'most significant performance improvements occur when KAN is applied across all three components,' but Table 2 shows 'Style Teacher KAN' achieving the highest Top-1 (66.49% vs. 66.26% for All Heads KAN). The paper should either identify the style-teacher placement as the best configuration or explain why the all-heads variant is preferred despite its lower Top-1 accuracy.
minor comments (6)
  1. [References [5] and [6]] References [5] and [6] appear to be placeholder entries ('Doe, A., Nguyen, L., Patel, R.' and 'Doe, J., Smith, J.') and should be replaced with real citations.
  2. [Reference list] Several references are duplicated in the list, e.g., [3]/[4] (SimCLR), [15]/[16] (MoCo), and [24]/[25] (KAN); these should be consolidated.
  3. [§4.2] §4.2 mentions training on an 'NVIDIA Quadro 4500 GPU,' which is likely a typo for a different GPU model; please correct the hardware description.
  4. [Eq. (1)] Eq. (1) does not define the output dimension d or the summation limits; please clarify how the KAN layer maps the input feature vector to the projected embedding.
  5. [Figure 3] Figure 3 is referenced in the text but no actual figure content appears; include the figure or remove the reference.
  6. [Dataset naming] The paper uses both 'Pandora 18K' and 'Pandora18k'; please standardize the dataset name.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the KAN-versus-MLP comparison is an external empirical evaluation, not a derivation from fitted inputs or a self-citation chain.

full rationale

The paper's central claim is that replacing MLP projection heads with KAN heads in a dual-teacher self-supervised distillation framework improves Top-1 accuracy on WikiArt and Pandora18k. That claim is evaluated by direct comparison against an external baseline (the dual-teacher framework of Luo et al.), with results reported as measured test-set accuracies in Tables 1 and 2. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no cited result is doing load-bearing work that reduces to the present claim; the KAN architecture is specified independently and the comparison is an empirical measurement. Hyperparameters and loss weights are chosen empirically, but that is standard experimental practice and does not make the evaluation circular. The manuscript does have evidentiary weaknesses, such as internal inconsistencies about whether all branches or only one teacher uses a KAN head (Section 3.1 and Figure 1 describe all branches as KAN, while Section 4 describes one teacher with an MLP head and one with a KAN head), the fact that the 'All Heads KAN (Ours)' row in Table 2 is not the highest-scoring row, and the absence of error bars or significance tests for single-run comparisons with deltas as small as 0.03 percentage points. These are correctness and reproducibility concerns about whether the improvement is real, not circularity: the result is not equivalent to its inputs by construction, and no self-citation loop or definitional reduction is present. Accordingly, the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim is empirical and rests on the standard assumption that KAN spline layers can improve feature projections, plus several domain assumptions about Gram matrices, augmentation, and EMA updates inherited from prior SSL work. No new theoretical entities are introduced. Key hyperparameters are unreported, which weakens both reproducibility and the fairness of the comparison.

free parameters (3)
  • KAN regularization weights lambda_L1 and lambda_smooth = not reported
    Eq. (5) uses these weights for the L1 sparsity and smoothness losses; values are never given, and they control how strongly the spline head is constrained.
  • segment deactivation rate for Ldeact = not reported
    The deactivation loss randomly turns off spline segments, but no probability or schedule is specified.
  • per-dataset hyperparameters (batch size, learning rate, input resolution) = batch 32/16, lr 0.0075/0.001, resolution 480/352 for WikiArt/Pandora18k
    Section 4.2 says these were empirically determined for optimal performance; without a separate tuning budget for each variant, the comparison may favor one side.
assumptions (5)
  • standard math Kolmogorov-Arnold representation theorem justifies a two-level spline composition as a useful approximator for continuous multivariate functions
    Invoked in Section 1 and Eq. (1) to justify KAN projection; the theorem is true, but its practical benefit in a shallow projection head is not proven.
  • domain assumption Gram matrices capture texture/style information and cosine similarity in Frobenius space aligns style
    Eq. (4) uses Gram matrices to enforce style alignment; this is inherited from prior art-style SSL work and not validated in this paper.
  • domain assumption Asymmetric augmentation (weak views for teachers, strong view for student) creates useful invariance
    Section 3.1 describes this strategy; it is standard in SSL but remains an assumption about the art domain.
  • domain assumption EMA update of teachers with momentum 0.99 prevents representation collapse and gives stable guidance
    Section 3.1, Eq. (7); inherited from BYOL/SimSiam literature, not re-derived here.
  • domain assumption Linear probe accuracy on frozen features measures representation quality
    Section 4 says they freeze the backbone and apply linear evaluation; this protocol assumes linear separability is the right yardstick.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Beyond Linear Bottlenecks: Spline-Based Knowledge Distillation for Culturally Diverse Art Style Classification." pith.science (2026). https://pith.science/paper/BPGCXKO4

@misc{pith2026250723436,
  author       = {Pith},
  title        = {Pith review of: Beyond Linear Bottlenecks: Spline-Based Knowledge Distillation for Culturally Diverse Art Style Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BPGCXKO4}},
  note         = {Machine review of arXiv:2507.23436}
}
read the original abstract

Art style classification remains a formidable challenge in computational aesthetics due to the scarcity of expertly labeled datasets and the intricate, often nonlinear interplay of stylistic elements. While recent dual-teacher self-supervised frameworks reduce reliance on labeled data, their linear projection layers and localized focus struggle to model global compositional context and complex style-feature interactions. We enhance the dual-teacher knowledge distillation framework to address these limitations by replacing conventional MLP projection and prediction heads with Kolmogorov-Arnold Networks (KANs). Our approach retains complementary guidance from two teacher networks, one emphasizing localized texture and brushstroke patterns, the other capturing broader stylistic hierarchies while leveraging KANs' spline-based activations to model nonlinear feature correlations with mathematical precision. Experiments on WikiArt and Pandora18k demonstrate that our approach outperforms the base dual teacher architecture in Top-1 accuracy. Our findings highlight the importance of KANs in disentangling complex style manifolds, leading to better linear probe accuracy than MLP projections.

Figures

Figures reproduced from arXiv: 2507.23436 by the authors.

Figure 1
Figure 1. Overall architecture of our proposed network. For each input image, three aug￾mented views X1, X2, and X3 are generated. X1 and X3 are fed to the two teacher networks, respectively. And X2 is input to the student network. Two teachers collabo￾ratively guide the student model, to ensure alignment with the teachers’ guidance. limits non-linear feature interactions, a critical constraint for modeling complex artistic s… view at source ↗
Figure 2
Figure 2. Class-wise confusion matrices for ConvNeXt-Base with KAN on (a) the Pan￾dora18k and (b) WikiArt test sets. Strong diagonal dominance and minimal cross-style errors highlight the model’s robustness. (a) Predictions from ConvNeXt-Base model (b) Predictions from ConvNeXt-Base + KAN model [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Comparative prediction examples from different models on the Pandora18k dataset. (a) Results from the convolutional baseline model, (b) Results from the ConvNeXt-Base architecture with KAN. Our framework significantly improves cross-dataset generalization across estab￾lished benchmarks like WikiArt and Pandora18k, outperforming previous dual￾teacher approaches through adaptive nonlinear projection [26]. Analysis of … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 29 canonical work pages

  1. [1]

    Expert Systems with Applications pp

    Cetinic, E., Lajic, T., Grgic, S.: Fine-tuning convolutional neural networks for fine art classification. Expert Systems with Applications pp. 107–118 (2018) 2 Beyond Linear Bottlenecks: Spline-Based Knowledge Distillation 11

    work page 2018

  2. [2]

    The Hubbard model on the honeycomb lattice: from static and dynamical mean-field theories to lattice quantum Monte Carlo simulations

    Cetinic, E., Lipic, T., Grgic, S.: Towards interpreting deep learning-based art style classification. arXiv preprint arXiv:1908.04307 (2019) 3

    work page Pith review arXiv 2019

  3. [3]

    1597–1607 (2020) 2

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations pp. 1597–1607 (2020) 2

    work page 2020

  4. [4]

    In: ICML

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: ICML. pp. 1597–1607 (2020) 3

    work page 2020

  5. [5]

    In: Proceedings of the International Conference on Machine Learning

    Doe, A., Nguyen, L., Patel, R.: Kolmogorov–arnold networks: Learning nonlinear projection through spline-based activation. In: Proceedings of the International Conference on Machine Learning. pp. 1023–1032 (2024) 3

    work page 2024

  6. [6]

    In: IEEE Transactions on Pattern Analysis and Machine In- telligence

    Doe, J., Smith, J.: Relational consistency via gram matrix alignment for self- supervised learning. In: IEEE Transactions on Pattern Analysis and Machine In- telligence. vol. 45, pp. 234–245 (2023) 3

    work page 2023

  7. [7]

    arXiv preprint arXiv:2010.11929 (2020) 8

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., et al.: An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929 (2020) 8

    arXiv 2020

  8. [8]

    In: AAAI Conference on Artificial Intelligence (2018) 3

    Elgammal, A., Liu, B., Elhoseiny, M., Mazzone, M.: The shape of art history in the eyes of the machine. In: AAAI Conference on Artificial Intelligence (2018) 3

    work page 2018

Show all 33 references
  1. [9]

    Expert Systems with Applications pp

    Falomir, Z., Museros, L., Sanz, I., Gonzalez-Abril, L.: Categorizing paintings in art styles based on qualitative color descriptors, quantitative global features and machine learning (qart-learn). Expert Systems with Applications pp. 83–94 (2018) 2

    2018

  2. [10]

    In: 2017 IEEE Winter Conference on Applications of Computer Vision (WACV)

    Florea, C., Toca, C., Gieseke, F.: Artistic movement recognition by boosted fusion of color structure and topographic description. In: 2017 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 569–577 (2017).https://doi. org/10.1109/WACV.2017.69 7

    2017 doi

  3. [12]

    In: Interspeech

    Fukuda, T., Suzuki, M., Kurata, G., Thomas, S., Cui, J., Ramabhadran, B.: Ef- ficient knowledge distillation from an ensemble of teachers. In: Interspeech. pp. 3697–3701 (2017) 3

    2017

  4. [13]

    In: IEEE/CVF Winter Conference on Applications of Computer Vision

    Gairola, S., Shah, R., Narayanan, P.: Unsupervised image style embeddings for retrieval and recognition tasks. In: IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2381–2399 (2020) 2

    2020

  5. [14]

    Garg, S., Jain, D.: Self-labeling refinement for robust representation learning with bootstrap your own latent (2022),https://arxiv.org/abs/2204.04545 3

    2022 arXiv

  6. [15]

    9729–9738 (2020) 2

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning pp. 9729–9738 (2020) 2

    2020

  7. [16]

    In: CVPR

    He, K., Fan, H., Wu, Y., Xie, S., Girshick, R.: Momentum contrast for unsupervised visual representation learning. In: CVPR. pp. 9729–9738 (2020) 3

    2020

  8. [17]

    In: Neural Information Processing Systems Deep Learning and Representation Learn- ing Workshop (2015) 3

    Hinton, G., Vinyals, O., Dean, J.: Distilling the knowledge in a neural network. In: Neural Information Processing Systems Deep Learning and Representation Learn- ing Workshop (2015) 3

    2015

  9. [18]

    283–285 (2017) 1

    Johnson, J.: Neural style representations and the large-scale classification of artistic style pp. 283–285 (2017) 1

    2017

  10. [19]

    In: British Machine Vision Conference (2013) 3

    Karayev, S., Hertzmann, A., Winnemoeller, H., Agarwala, A., Darrell, T.: Recog- nizing image style. In: British Machine Vision Conference (2013) 3

    2013

  11. [20]

    In: British Machine Vision Conference (BMVC)

    Karayev, S., Trentacoste, M., Han, H., Agarwala, A., Darrell, T., Hertzmann, A., Winnemoeller, H.: Recognizing image style. In: British Machine Vision Conference (BMVC). Nottingham, UK (September 2014), https://bmva-archive.org.uk/ bmvc/2014/files/paper121/index.html 7 12 AZ. ...

    2014

  12. [21]

    arXiv preprint arXiv:1311.3715 (2013) 2

    Karayev, S., et al.: Recognizing image style. arXiv preprint arXiv:1311.3715 (2013) 2

    2013 arXiv

  13. [22]

    In: Advances in Neural Information Processing Sys- tems

    Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep con- volutional neural networks. In: Advances in Neural Information Processing Sys- tems. pp. 1097–1105 (2012) 3

    2012

  14. [23]

    arXiv preprint arXiv:2201.03545 (2022) 8

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A convnet for the 2020s. arXiv preprint arXiv:2201.03545 (2022) 8

    2022 arXiv

  15. [25]

    arXiv preprint arXiv:2404.19756 (2024) 3

    Liu, Z., Wang, Y., Vaidya, S., Ruehle, F., Halverson, J., Soljačić, M., Hou, T.Y., Tegmark, M.: Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756 (2024) 3

    2024 arXiv

  16. [26]

    Applied Soft Computing174, 112964 (2025) 2, 10

    Luo, M., Liu, L., Lu, Y., Suen, C.Y.: Art style classification via self-supervised dual-teacher knowledge distillation. Applied Soft Computing174, 112964 (2025) 2, 10

    2025

  17. [27]

    Applied Soft Computing174, 112964 (2025) 3, 4

    Luo, M., Liu, L., Lu, Y., Suen, C.Y.: Art style classification via self-supervised dual-teacher knowledge distillation. Applied Soft Computing174, 112964 (2025) 3, 4

    2025

  18. [28]

    Peterson, J.C., Soulos, P., Nematzadeh, A., Griffiths, T.L.: Learning hierarchical visual representations in deep neural networks using hierarchical linguistic labels (2018), https://arxiv.org/abs/1805.07647 2

    2018 arXiv

  19. [29]

    In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Pham, C., Hoang, T., Do, T.T.: Collaborative multi-teacher knowledge distillation for learning low bit-width deep neural networks. In: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 6435–6443 (2023) 3

    2023

  20. [30]

    In: IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (2020) 3

    Pham, V., et al.: Multi-teacher multi-task knowledge distillation for semantic seg- mentation. In: IEEE/CVF Conference on Computer Vision and Pattern Recogni- tion (2020) 3

    2020

  21. [31]

    In: International Conference on Learning Representations (2015) 3

    Simonyan, K., Zisserman, A.: Very deep convolutional networks for large-scale im- age recognition. In: International Conference on Learning Representations (2015) 3

    2015

  22. [32]

    Tan, M., Le, Q.V.: Efficientnet: Rethinking model scaling for convolutional neural networks (2019) 8

    2019

  23. [33]

    In: Proceedings of the IEEE International Conference on Computer Vision

    Zhang, W., Li, M., Chen, J.: A multi-task self-supervised framework for art style classification. In: Proceedings of the IEEE International Conference on Computer Vision. pp. 1234–1242 (2023) 3

    2023

  24. [34]

    In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Zhang, Y., Chen, H., Wang, L., Yang, M.H., Li, W.: Three-element learning: A uni- fied framework for multi-task self-supervised representation learning. In: Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11234–11244 (2022) 2

    2022

  25. [35]

    Applied Sciences 14(24), 11717 (2024).https://doi.org/10.3390/app142411717 2

    Zhou, D., Zhou, D., Wei, G., Yuan, X.: Three-dimensional shape reconstruction from digital freehand design sketching based on deep learning techniques. Applied Sciences 14(24), 11717 (2024).https://doi.org/10.3390/app142411717 2

    2024 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.