Pith. sign in

REVIEW 4 major objections 6 minor 37 references

Neural network task specialization via domain constraining

T0 review · 4 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read Restricting a neural network's label space to a relevant subset improves its accuracy on that subset without adding data.

desk verdict Useful extraction trick and a clean semantic-vs-random comparison, but the central 'solely constraining' claim is confounded by the missing full-label fine-tuning control. read the letter →

arxiv 2504.19592 v1 pith:AQ6AW5N2 submitted 2025-04-28 cs.CV cs.AI

Roman Malashin , Daniil Ilyukhin This is my paper · ORCID
classification cs.CVcs.AI
keywords neuralnetworkspecializationdomainconstrainingdynamicconfigurabilityimageclassificationobjectdetectionfine-tuninglabelspacerestrictiontransferlearning
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 claims that a generalist image-analysis model can be turned into a specialist for a subset of classes and gain accuracy on that subset with no additional data and no change of training regime, solely by constraining the label space. The proposed procedure extracts the specialist by keeping only the rows of the final linear layer that correspond to the relevant classes, and then fine-tunes the network on images from those classes. On ImageNet classification with MobileNet V3 Small, specialists improve average accuracy by about 1.8 percentage points across five 200-class blocks; on COCO detection with YOLOv5, most superclass specialists improve fitness over the generalist. The paper also reports that semantically coherent subsets built from WordNet give larger gains than random subsets, and that in a larger detector, freezing some layers during specialization can improve performance. If the claim holds, it offers a cheap path to domain-specific accuracy from an existing generalist, without retraining on new data.

What carries the argument

The load-bearing mechanism is specialist extraction: composing a new classification head $W_S$ and $b_S$ by selecting rows of the final linear layer belonging to relevant labels $S$, followed by fine-tuning on the constrained data set $D_S$. Extraction makes the loss depend only on relevant classes, so the network no longer spends capacity suppressing probabilities of irrelevant classes and the generalist's features are preserved. For detectors, where the head has objectness scores and no softmax, the paper approximates the same forgetting by suppressing the outputs of irrelevant classes during training. A gradient analysis of the last layer supports the claim that training only the head with frozen features cannot guarantee complete forgetting, which is why extraction before full fine-tuning is proposed.

What would settle it

Run the same fine-tuning schedule used for specialists, Adam at learning rate $10^{-4}$ for 200 epochs, on the generalist while keeping the full 1000-class label space, then measure per-subset accuracy. If the full-label fine-tune matches the extracted specialists' gains, the effect is continued training rather than domain constraining. A sharper version compares extracted-head fine-tuning against randomly reinitialized-head fine-tuning on the same subset with identical optimization.

Watch

Extended reading notes

Core claim

The central discovery is that excluding irrelevant classes from both the label space and the training data, rather than adding data or changing architecture, can improve a network's accuracy on the remaining classes. The key to doing this is what the paper calls efficient forgetting: replacing the final linear layer with the submatrix $W_S = W[S,:]$ and bias $b_S = b[S]$, selecting only the rows belonging to the relevant label set $S$, before fine-tuning. This extraction preserves the features learned by the generalist while eliminating the logits of irrelevant classes, so fine-tuning can reallocate capacity toward the relevant domain. After fine-tuning, the extracted specialist outperforms the generalist on the relevant subset, and the paper's UMAP and CKA analyses indicate that specialization mainly reshapes the top layers while preserving lower-level features. The paper further argues that specialization is only effective when the constrained class subset is semantically coherent, and that gradually constraining the domain, for example from all COCO classes to animals to horses, can give larger gains than direct specialization.

Load-bearing premise

The gains are attributed to excluding irrelevant classes, but the paper never fine-tunes the generalist on the full label space with the same optimizer, learning rate, and epoch count, so the improvements could come from additional training on the relevant data alone.

Editorial extensions

If this is right

  • A generalist can act as a router and refine its own prediction with extracted specialists: on ImageNet the dynamically configured system raises top-1 accuracy from 67.5% to 69.1%.
  • Semantic coherence is what makes specialization pay off: WordNet-based specialists gain 1.5 to 2.8 percentage points, while random subsets gain at most 0.5 percentage points.
  • Gradual domain constraining helps in detection: a horse specialist initialized from an animal specialist reaches fitness 0.936, versus 0.917 when initialized directly from the generalist.
  • Layer-wise modularity appears in larger models: freezing parts of the neck or head of YOLOv5x during specialization improves fitness over full fine-tuning.
  • No extra data or new training regime is needed for the core effect, making specialization a practical second training phase for an already trained model.

Reading between the lines

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

  • The extraction trick should transfer to any model with a linear output layer, including multi-label or regression heads, where irrelevant classes become irrelevant output dimensions; the paper does not test this.
  • The results suggest that some test-time errors of a generalist come from competition among irrelevant classes, so a router that selects a specialist by predicted class should inherit most of the gains; the paper's dynamic system result supports this.
  • Specialization could serve as a lightweight continual-learning primitive, where a generalist stays frozen and new domains are handled by newly extracted specialists; the paper does not evaluate this scenario.
  • If the confound identified below is resolved and the gains remain, the mechanism implies that shared feature space across many classes can hurt a narrow task, which could be tested by measuring logit competition before and after extraction.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 a specialization procedure for neural network classifiers and detectors: starting from a generalist trained on the full label space, restrict the label set to a task-specific subset S, extract the corresponding rows of the final linear layer, and fine-tune on the constrained subset. The authors report accuracy gains on ImageNet classification (MobileNet V3 Small) and COCO object detection (YOLOv5m/x) and attribute these gains solely to constraining the class label space, without additional data or altered training regimes. They also compare semantically coherent and random class groupings, analyze feature-space evolution with UMAP and CKA, and examine layer freezing for detector specialists.

Significance. If the causal claim were established, the paper would provide a simple and practical way to improve accuracy in constrained deployment scenarios, and the extraction-before-tuning procedure is a useful technical contribution. Strengths of the study include the use of a random-class control (Table 3), experiments on two tasks, and the layer-freezing analysis in Table 7. However, the central claim is not isolated from the effect of additional fine-tuning, because the generalist and specialist training protocols differ in optimizer, epoch budget, and data selection; the reported gains are also single-run results without error bars. With the missing control experiments added, the paper could make a solid contribution; as it stands, the evidence is suggestive rather than conclusive.

major comments (4)
  1. [Section 3, Training parameters; Table 1] The central claim that specialization improves accuracy 'solely by constraining class label space' is not isolated from continued fine-tuning. Generalists are trained for 600 epochs with RMSprop, while specialists are extracted and then fine-tuned for 200 epochs with Adam. No control is reported in which the generalist is fine-tuned on the full 1000-class label space with the same Adam settings and epoch budget. Without such a control, the +1.8% average gain in Table 1 is compatible with the explanation that additional training on in-domain data with a different optimizer improves accuracy, rather than exclusion of irrelevant classes. This also contradicts the abstract's phrase 'without changing training regimes,' since the optimizer and epoch count are changed.
  2. [Section 3, Table 3] The random-specialist comparison is a good control for semantic coherence, but it does not control for the effect of fine-tuning itself. Random specialists also receive 200 epochs of Adam training on their subsets; the fact that they gain only about 0.0-0.5% shows that the gain depends on the subset, not that the gain is caused by constraining rather than by additional training. The missing control remains a full-label-space fine-tune with identical optimizer and epoch settings; without it, the contrast only establishes that semantic subsets benefit more from this particular fine-tuning procedure.
  3. [Section 4 (detector experiments); Tables 4-6] The detector experiments have the same confounding structure. The generalist baseline is trained from random initialization on all 80 COCO classes, while specialists are initialized from the generalist and fine-tuned for 300 epochs with dataset-specific settings. There is no control in which the generalist is fine-tuned on the full COCO label space with the same settings, so the reported ΔF improvements in Tables 4 and 6 could reflect additional training rather than domain constraining. Please add that control or a comparable whole-dataset fine-tune before claiming that specialization is the cause of the gains.
  4. [All experimental sections (Tables 1-7)] All numerical comparisons are based on single runs without standard errors or confidence intervals. Given the small effect sizes (for example, +1.8% in Table 1 and +0.003 for Food in Table 6), single runs do not establish that the reported differences are reliable. Report results over multiple seeds or provide statistical significance measures.
minor comments (6)
  1. [Section 4, Training parameters] The optimizer, learning rate, and learning-rate schedule for detector fine-tuning are not specified; only epoch count, batch size, and augmentation settings are given. Please provide full hyperparameters for reproducibility.
  2. [Metrics section] The formula 'mAP = TP / (TP + FN + FP)' is not the standard definition of mean average precision. Since the paper uses the YOLO fitness metric, either define mAP correctly or state that this expression is the YOLO fitness approximation.
  3. [Appendix A] The gradient notation is inconsistent. The text derives ∂L/∂w_ij = p_j h_i, but then writes Δw_ij ∝ -p_i h_j, swapping indices. Please correct the subscripts.
  4. [Table 5] The header contains the typo 'Initializa-toion', and decimal commas (0,9) are used in Table 5 while decimal points are used elsewhere. Use a consistent decimal separator.
  5. [Section 3, Experiment 2] The semantic grouping procedure is deferred to reference [14], an unpublished preprint by the authors. Please describe the WordNet grouping in sufficient detail in the paper so that the experiments can be reproduced without consulting [14].
  6. [Introduction] The Introduction says 'Section 3 extends the approach to object detection,' but the detector experiments appear in an unnumbered section after Section 3. Please renumber the sections.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the specialization gains are empirical measurements against external baselines, not derivations from fitted inputs or self-citations.

full rationale

The paper's central claim is that constraining the label space and fine-tuning (after extracting the relevant rows of the final linear layer) improves accuracy on that subspace. This claim is tested directly: the generalist and specialists are evaluated on the same validation subsets (Table 1), semantic versus random partitions are compared (Table 3), and detector specialists are compared with the generalist on the same COCO domains (Tables 4-6). The label subsets come from external structure (ImageNet synset ordering, WordNet, COCO superclasses) or from random partitions, not from the outcome data, so no prediction is equivalent to a fitted input. The extraction operation W_S = W[S,:] is a mathematical identity with evaluating the generalist while ignoring irrelevant classes; the paper uses it as a baseline, not as the source of the claimed gains. Self-citations, notably [14] for the WordNet grouping procedure and hypernym-bias motivation, are reproducibility pointers and motivational context; Table 3's random-vs-semantic comparison and the accuracy measurements are made in this paper, so the central result does not reduce to the authors' prior work. The main weakness is experimental rather than circular: specialists are fine-tuned with Adam for 200 epochs while the generalist was trained with RMSprop for 600 epochs, and no full-label-space fine-tuning control is reported; this confound affects causal attribution but is not a case of a quantity being defined or fitted in terms of the claimed outcome.

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

The paper's central claim is empirical and introduces no fitted constants or invented entities. It relies on standard softmax classifier theory and on two domain assumptions: the generalist is not fully converged, and WordNet's semantic groupings match the model's error structure. The latter is supported mainly by the authors' own prior work [14].

assumptions (3)
  • domain assumption The generalist model trained on the full dataset does not reach zero training loss, which is stated as necessary for specialization gains.
    Section 3, model paragraph: 'it is necessary that generalist model trained on the whole dataset does not reach zero training loss, which is satisfied for parameter efficient models.' This is an unproved background condition for the effect.
  • domain assumption WordNet hypernym structure aligns with the network's learned feature manifold and error structure.
    Section 3, Experiment 2 uses WordNet groups for semantic specialists and cites the authors' own prior work [14] for the grouping procedure. If this alignment fails, the semantic-specialist advantage disappears.
  • standard math The last layer of the classifier is linear and the feature vector h is treated as fixed for the gradient analysis in Appendix A.
    Appendix A derives gradients under z = Wh + b and assumes lower-layer features are frozen. This is a standard modeling assumption for analyzing the final layer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural network task specialization via domain constraining." pith.science (2026). https://pith.science/paper/AQ6AW5N2

@misc{pith2026250419592,
  author       = {Pith},
  title        = {Pith review of: Neural network task specialization via domain constraining},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AQ6AW5N2}},
  note         = {Machine review of arXiv:2504.19592}
}
read the original abstract

This paper introduces a concept of neural network specialization via task-specific domain constraining, aimed at enhancing network performance on data subspace in which the network operates. The study presents experiments on training specialists for image classification and object detection tasks. The results demonstrate that specialization can enhance a generalist's accuracy even without additional data or changing training regimes: solely by constraining class label space in which the network performs. Theoretical and experimental analyses indicate that effective specialization requires modifying traditional fine-tuning methods and constraining data space to semantically coherent subsets. The specialist extraction phase before tuning the network is proposed for maximal performance gains. We also provide analysis of the evolution of the feature space during specialization. This study paves way to future research for developing more advanced dynamically configurable image analysis systems, where computations depend on the specific input. Additionally, the proposed methods can help improve system performance in scenarios where certain data domains should be excluded from consideration of the generalist network.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 18 canonical work pages

  1. [1]

    Hinton, O

    G. Hinton, O. Vinyals, J. Dean, Distilling the knowledge in a neural network, arXiv preprint arXiv:1503.02531 (2015). doi:10.48550/ arXiv.1503.02531

    arXiv doi:10.48550/arxiv.1503.02531 2015

  2. [2]

    Xiao B. et al. Florence-2: Advancing a unified representation for a variety of vision tasks //Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. –

    work page

  3. [3]

    Kirillov A. et al. Segment anything //Proceedings of the IEEE/CVF international conference on computer vision. –

    work page

  4. [4]

    Wang X. et al. Images speak in images: A generalist painter for in-context visual learning //Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. – 2023. – С. 6830-6839

    work page 2023

  5. [5]

    Biedenkapp, H

    A. Biedenkapp, H. F. Bozkurt, T. Eimer, F. Hutter, M. Lindauer, Dynamic algorithm configuration: Foundation of a new meta -algorithmic framework, in: Proceedings of the Twenty-fourth European Conference on Artificial Intelligence, Vol. 163, 2020. doi:10.3233/FAIA200122

    doi:10.3233/faia200122 2020

  6. [6]

    Shazeer, A

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, J. Dean, Outrageously large neural networks: The sparsely-gated mixture -of-experts layer, arXiv preprint arXiv:1701.06538 (2017). doi: 10.48550/arXiv.1701.06538

    arXiv doi:10.48550/arxiv.1701.06538 2017

  7. [7]

    Agarwal, A

    N. Agarwal, A. Sondhi, K. Chopra, G. Singh, Transfer learning: Survey and classification, Smart Innovations in Communica- tion and Computational Sciences: Proceedings of ICSICCS 2020 (2021) 145– 155doi:10.1007/978-981-15-5345-5_13

    work page doi:10.1007/978-981-15-5345-5_13 2021

  8. [8]

    B. Zhou, N. Kalra, P. Krähenbühl, Domain adaptation through task distillation, in: European Conference on Computer Vision, Springer, 2020, pp. 664–680. doi:10.1007/978-3-030-58574- 7_40

    work page doi:10.1007/978-3-030-58574- 2020

Show all 37 references
  1. [9]

    Wang, J.-Y

    T. Wang, J.-Y. Zhu, A. Torralba, A. A. Efros, Dataset distilla- tion, arXiv preprint arXiv:1811.10959 (2018). doi:10.48550/arXiv.1811.10959. preprint 8 Under review at Computer Optics

    2018 doi arXiv

  2. [10]

    Malashin, Extraction of object hierarchy data from trained deep- learning neural networks via analysis of the confusion matrix, J

    R. Malashin, Extraction of object hierarchy data from trained deep- learning neural networks via analysis of the confusion matrix, J. Opt. Tech. 83 (2016) 599 –603. doi:10.1364/jot.83.000599

    2016 doi

  3. [11]

    Stringer C. et al. Cellpose: a generalist algorithm for cellular segmentation //Nature methods. – 2021. – Т. 18. – №. 1. – С. 100-106

    2021

  4. [12]

    Arpit, S

    D. Arpit, S. Jastrze˛bski, N. Ballas, D. Krueger, E. Bengio, M. S. Kanwal, T. Maharaj, A. Fischer, A. Courville, Y. Bengio, et al., A closer look at memorization in deep networks, in: Inter- national conference on machine learning, PMLR, 2017, pp. 233–242. doi:10.5555/3305381.3305406

    2017

  5. [13]

    A. M. Saxe, J. L. McClelland, S. Ganguli, A mathematical the- ory of semantic development in deep neural networks, Proceed- ings of the National Academy of Sciences 116 (23) (2019) 11537–11546.10.1073/pnas.1820226116

    2019 doi

  6. [14]

    Hypernym Bias: Unrav- eling Deep Classifier Training Dynamics through the Lens of Class Hierarchy //arXiv preprint arXiv:2502.12125

    Malashin R., Yachnaya V., Mullin A. Hypernym Bias: Unrav- eling Deep Classifier Training Dynamics through the Lens of Class Hierarchy //arXiv preprint arXiv:2502.12125. – 2025

    2025 arXiv

  7. [15]

    J. A. Mendez, E. Eaton, How to reuse and compose knowledge for a lifetime of tasks: A survey on continual learning and func- tional composition, arXiv preprint arXiv:2207.07730 (2022). doi:10.3233/FAIA200122

    2022 arXiv

  8. [16]

    Bengio, P.-L

    E. Bengio, P.-L. Bacon, J. Pineau, D. Precup, Conditional com- putation in neural networks for faster models, arXiv preprint arXiv:1511.06297 (2015). doi:10.48550/arXiv.1511.06297

    2015 doi arXiv

  9. [17]

    R. A. Jacobs, M. I. Jordan, S. J. Nowlan, G. E. Hinton, Adaptive mixtures of local experts, Neural computation 3 (1) (1991) 79–

    1991

  10. [18]

    Fedus, B

    W. Fedus, B. Zoph, N. Shazeer, Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity, Journal of Machine Learning Research 23 (120) (2022) 1–39. doi:10.48550/arXiv.2101.03961

    2022 doi arXiv

  11. [19]

    Z. Qiu, Z. Huang, B. Zheng, K. Wen, Z. Wang, R. Men, I. Ti- tov, D. Liu, J. Zhou, J. Lin, Demons in the detail: On imple- menting load balancing loss for training specialized mixture-of- expert models, arXiv preprint arXiv:2501.11873 (2025). doi:10.48550/arXiv.2501.11873

    2025 doi arXiv

  12. [20]

    Bolukbasi, J

    T. Bolukbasi, J. Wang, O. Dekel, V. Saligrama, Adaptive neu- ral networks for efficient inference, in: International Confer- ence on Machine Learning, PMLR, 2017, pp. 527 –536. doi:10.48550/arXiv.1702. 07811

    2017 doi

  13. [21]

    R. O. Malashin. Sparsely ensembled convolutional neural net- work classifiers via reinforcement learning, in: ICMLT 2021: 6th International Conference on Machine Learning Technolo- gies, Jeju Island, Republic of Korea, April 23 - 25, 2021, 2021, pp. 102 –110. doi:10.1145/3468 ...

    2021

  14. [22]

    R. O. Malashin, Principle of least action in dynamically config- ured image analysis systems, Journal of Optical Technology 86 (11) (2019) 678. doi:10.1364/jot.86.000678. URL http://dx.doi.org/10.1364/jot.86.000678

    2019 doi

  15. [24]

    Shelepin, N

    Y. Shelepin, N. Krasilnikov, G. Trufanov, A. Harauzov, S. Pro- nin, A. Foking, The principle of least action and viusal perception, in: Twenty-ninth European Conference on Visual Perception, 2006. doi: 10.1068/p5366

    2006 doi

  16. [25]

    Zhuge, H

    M. Zhuge, H. Liu, F. Faccio, D. R. Ashley, R. Csordás, A. Go- palakrishnan, A. Hamdi, H. A. A. K. Hammoud, V. Herrmann, K. Irie, et al., Mindstorms in natural language-based societies of mind, arXiv preprint arXiv:2305.17066 (2023). doi:10.48550/arXiv.2305.17066

    2023 doi

  17. [26]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al., Imagenet large scale visual recognition challenge, International journal of computer vision 115 (2015) 211 –252. doi:10.1007/s11263-015-0816-y

    2015 doi

  18. [27]

    Howard, M

    A. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, M. Tan, W. Wang, Y. Zhu, R. Pang, V. Vasudevan, et al., Searching for mobilenetv3, in: Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 1314 –1324. doi:10.1109/ICCV.2019.00140

    2019

  19. [28]

    Hinton, N

    G. Hinton, N. Srivastava, K. Swersky, Neural networks for ma- chine learning lecture 6a overview of mini-batch gradient de- scent, Cited on 14 (8) (2012) 2

    2012

  20. [29]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimiza- tion, arXiv preprint arXiv:1412.6980 (2014). doi:10.48550/arXiv.1412.6980

    2014 doi arXiv

  21. [30]

    Fellbaum, Wordnet, in: Theory and applications of ontology: computer applications, Springer, 2010, pp

    C. Fellbaum, Wordnet, in: Theory and applications of ontology: computer applications, Springer, 2010, pp. 231 –243. doi:10.1007/ 978-90-481-8847-5_10

    2010

  22. [31]

    Umap: Uniform manifold ap- proximation and projection for dimension reduction //arXiv preprint arXiv:1802.03426

    McInnes L., Healy J., Melville J. Umap: Uniform manifold ap- proximation and projection for dimension reduction //arXiv preprint arXiv:1802.03426. – 2018

    2018 arXiv

  23. [32]

    Kornblith S. et al. Similarity of neural network representations revisited // International conference on machine learning. – PMLR, 2019. – P. 3519 -3529. doi: 10.48550/arXiv.1905.00414

    2019 doi arXiv

  24. [33]

    Jocher, et al., ultralytics/yolov5: v3.0 - third release, 3, 6 (De- cember 2020)

    G. Jocher, et al., ultralytics/yolov5: v3.0 - third release, 3, 6 (De- cember 2020). URL: https://github.com/ultralytics/yolov5

    2020

  25. [34]

    T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ra- manan, P. Dollár, C. L. Zitnick, Microsoft coco: Common ob- jects in context, in: Proceedings of the European Conference on Computer Vision (ECCV), Vol. 8693 of Lecture Notes in Com- puter Science, Springer, 2014, pp...

    2014 doi

  26. [35]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. Williams, J. Winn, A. Zis- serman, The pascal visual object classes (voc) challenge, Inter- national journal of computer vision 88 (2010) 303 –338. doi:10.1007/s11263-009-0275-4

    2010 doi

  27. [36]

    Y. Cui, Y. Song, C. Sun, A. Howard, S. Belongie, Large scale fine-grained categorization and domain-specific transfer learn- ing, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4109 –4118. doi:10.1109/CVPR.2018.00432

    2018

  28. [37]

    Structural analysis of images of 3-D scenes

    L. Bossard, M. Guillaumin, L. Van Gool, Food-101–mining discriminative components with random forests, in: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Swit- zerland, September 6 -12, 2014, Proceedings, Part VI 13, Springer, 2014, pp. 446–461. doi:10.1007/978-3...

    2014 doi

  29. [87]

    doi:10.1162/neco.1991.3.1.79

    1991 doi

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.