Pith. sign in

REVIEW 6 major objections 6 minor 52 references

Proto-FG3D: Prototype-based Interpretable Fine-Grained 3D Shape Classification

T0 review · 6 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Proto-FG3D replaces the softmax classifier with non-parametric prototype matching for fine-grained 3D shape classification and reports state-of-the-art accuracy with built-in interpretability.

desk verdict A sensible prototype-based framing for fine-grained 3D classification, but Eq. (6) as written destroys prototype diversity; this needs a fix and re-verification before the results can be trusted. read the letter →

arxiv 2505.17666 v1 pith:UHUXKIAM submitted 2025-05-23 cs.CV

classification cs.CV
keywords fine-grained3Dshapeclassificationprototypelearninginterpretablemachineoptimaltransportmulti-viewclassimbalancecase-basedreasoning
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

Proto-FG3D claims that fine-grained 3D shape classification can be improved by replacing the parametric softmax classifier with non-parametric prototype learning. The framework groups multi-view features into learnable class-specific prototypes using an entropy-regularized optimal transport assignment, refines them with online momentum updates, and decides by nearest-prototype retrieval. The authors report consistent gains in accuracy over CNN- and ViT-based state-of-the-art baselines on the FG3D benchmark and ModelNet40, and a narrower gap between average instance accuracy and average class accuracy, which they attribute to better handling of severe subclass imbalance. The resulting predictions are ad-hoc interpretable because each decision is explained by the matched prototype, visualized as a representative shape and view-level activation map.

What carries the argument

The central machinery is the prototype pool coupled with three mechanisms. Prototype Association formulates view-to-prototype assignment as an entropy-regularized optimal transport problem (Eqs. 1-4), solved with Sinkhorn-Knopp iteration, with constraints meant to enforce uniqueness (each feature assigned to exactly one prototype) and equipartition (balanced assignment across prototypes). Online Prototype Update refines each prototype by an exponential moving average of the features assigned to it, with a momentum schedule that starts near 0.999. Prototype-based Supervised Representation Learning combines a cross-entropy loss on nearest-prototype distance (Eq. 8) with a view-prototype contrastive loss (Eq. 9). The decision rule that carries the argument is Eq. (10), nearest-prototype retrieval, which turns each prediction into transparent case-based reasoning.

What would settle it

Extract the Sinkhorn assignment matrix for a trained model and check whether it is nearly one-hot: if typical entries are far from 1/0 (e.g., many features have no prototype with assignment above 0.5), the claimed uniqueness constraint is violated in practice. A direct test is to round the soft assignments to hard one-hot labels during prototype updates and training; if accuracy or class balance metrics drop sharply, the method's behavior depends on the fractional relaxation rather than on the one-to-one clustering it describes.

Watch

Extended reading notes

Core claim

Proto-FG3D is a framework for fine-grained 3D shape classification built on a shared pool of K trainable prototypes per class. Multi-view features are softly assigned to prototypes by solving an entropy-regularized optimal transport problem under uniqueness and equipartition constraints, prototypes are updated online via exponential moving averages of their assigned feature centroids, and training combines a cross-entropy loss over feature-to-prototype distances with a prototype-view contrastive loss. Inference is non-parametric: a test view is classified by the nearest prototype in the pool. Experiments on the FG3D dataset (airplane, car, chair subcategories) and ModelNet40 show the framework improves both average instance accuracy and, in most configurations, average class accuracy over strong CNN and ViT baselines, and the prototypes provide transparent case-based explanations of each prediction.

Load-bearing premise

The load-bearing premise is that the entropy-regularized optimal transport solution, which produces fractional scores, faithfully realizes the paper's stated one-to-one assignment of each view to a single prototype; if that relaxation is unfaithful, the prototype updates and losses are not doing what the description claims.

Editorial extensions

If this is right

  • If Proto-FG3D's gains hold, non-parametric prototype classifiers become a viable substitute for softmax heads in 3D shape recognition, especially where interpretability is required.
  • The narrower gap between average instance accuracy and average class accuracy indicates that prototype-based allocation helps minority subclasses, addressing the severe class imbalance in fine-grained 3D benchmarks.
  • The framework can be bolted onto existing CNN and ViT backbones (MVCNNnew, GVCNN, SMVCNN, DAN, VSFormer) with consistent improvements, so current architectures can adopt it without redesign.
  • Because inference is nearest-prototype retrieval, predictions come with a human-readable explanation (the matched prototype) by construction, not via a post-hoc explanation tool.

Reading between the lines

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

  • If the soft optimal-transport relaxation is shown to be a faithful proxy for hard clustering, the same prototype-association mechanism could transfer to other multi-view tasks such as 3D part segmentation or retrieval, where balanced assignment and interpretability are also valuable.
  • The entropy-regularized OT formulation is essentially a clustering with a fairness constraint (equipartition); one testable extension is to weight the equipartition constraint by class prior so it actively counteracts subclass imbalance rather than just equalizing assignment counts.
  • The paper's interpretability evidence is qualitative (prototype images and CAM overlays); a quantitative test would be whether a human can predict the model's decision from the top-3 prototypes alone, which would elevate ad-hoc interpretability from demonstration to measurement.
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

6 major / 6 minor

Summary. The manuscript proposes Proto-FG3D, a prototype-based multi-view 3D shape classifier that replaces the parametric softmax head with a per-class pool of K prototypes. Features from rendered views are assigned to prototypes by entropy-regularized optimal transport (Sinkhorn/APDAGD), prototypes are updated online with momentum, and training combines a nearest-prototype cross-entropy loss with a view-prototype contrastive loss. Inference is nearest-prototype matching, providing case-based interpretability. The paper reports accuracy comparisons on FG3D and ModelNet40 with several CNN and ViT backbones, qualitative prototype/CAM visualizations, and ablations on the FG3D-Airplane subset.

Significance. If the mechanism worked as described, replacing softmax with OT-driven prototype matching in multi-view 3D recognition would be a relevant contribution: it offers a concrete route to ad-hoc interpretability and a non-parametric decision layer, and the paper tests it across multiple backbones and two benchmarks. The main strengths are the breadth of the backbone integrations and the attempt to connect prototype learning to fine-grained and class-imbalanced settings. However, the paper currently does not define a consistent algorithm: the prototype update equation, if read literally, cannot specialize prototypes, and the experimental reporting contains contradictions and lacks variance estimates. These issues block acceptance in the current form.

major comments (6)
  1. [Section 3.2, Eq. (6)] The prototype update as written averages the same class-wide view mean for every set S^c_k, because (1/V^c) sum_{i=1}^{V^c} h^c_i does not depend on membership in S^c_k. Consequently \bar H^c_k is identical for all k, and the EMA update in Eq. (5) cannot specialize prototypes to sub-clusters. If the intended operation is the mean of only the assigned features, the formula must be corrected and the experiments repeated; as written, the described method cannot produce the Table 1 results.
  2. [Section 4.2, Table 1] The statement that integrating Proto-FG3D 'consistently improves all baselines' is contradicted by several rows in Table 1: SMVCNN+Ours (A) Airplane ACA is 90.98 versus the 91.36 baseline, MVCNNnew+Ours (A) and (S) ModelNet ACA are 91.58 versus 91.73, DAN+Ours (A) Car ACA is 71.07 versus 72.78, and DAN+Ours (S) ModelNet AIA is 94.98 versus 95.02. The claim and the arrow annotations should be revised to acknowledge these degradations explicitly.
  3. [Section 4.2] The claim that 'MVCNNnew+Ours (S) achieves ACA surpassing AIA' is only true for Car (ACA 77.25 versus AIA 77.11); for Airplane ACA is 91.54 versus AIA 94.81, and for Chair ACA is 79.33 versus AIA 84.56. The sentence should be corrected or restricted to the Car category.
  4. [Section 4.4, Table 4] The reported 'best' numbers are inconsistent across sub-tables: the combined-loss row in (a) gives AIA 95.36/ACA 92.95, (b) with K=20 gives 95.36/92.87, and (c) with eta_0=0.999 and (d) with alpha=0.20 both give 95.63/93.95. If these are the same default configuration the numbers must match; otherwise the exact settings of each ablation row need to be stated.
  5. [Section 4.1-4.2] All results are single runs without multiple seeds, error bars, or significance tests. Several differences are very small, for example VSFormer+Ours (A) ModelNet AIA 95.30 versus 95.26 and ACA 92.89 versus 92.89, so the headline improvements cannot be distinguished from training noise without repeated runs and variance reporting.
  6. [Section 3.1, Eqs. (2)-(4)] The text says each feature is assigned to exactly one prototype, but the constraints are soft column-sum constraints and Eq. (4) produces fractional entries in Z^c. The manuscript should clarify whether the downstream losses in Eqs. (7)-(9) use hard or soft assignments and how this is reconciled with the 'exactly one' statement.
minor comments (6)
  1. [Section 3.1, Eq. (1)] The term -mean_{Z^c}{...} is not defined; the optimization objective is therefore ambiguous and should be rewritten.
  2. [Section 1 and Section 3] The method is repeatedly called 'non-parametric', but the prototypes in Q^c are trainable parameters. Please clarify that only the decision layer, not the prototype pool, is non-parametric.
  3. [Section 4.3, Table 2] The caption says 'top-3 prototypes from both majority and minority subclasses for each dataset', but the table shows only two datasets and three similarity scores per class; please clarify the displayed content.
  4. [Section 4.1] The statement that the method reduces parameters and FLOPs by 1% should be replaced with the specific measured numbers and a precise description of which modules are included in the comparison.
  5. [Section 3.3, Eq. (8)] The notation d(h_v, Q^c) is used before Q^c is defined as a set; please define the distance to a prototype pool explicitly.
  6. [References] Reference [50] contains a formatting error: 'Cuturi M Sinkhorn Distances. Lightspeed computation of optimal transport' should be corrected to the standard citation format.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark claims are externally measured and the prototype equations, despite a correctness concern in Eq. (6), do not reduce to fitted inputs.

full rationale

The central derivation chain is self-contained and externally benchmarked. Prototype association (Eqs. 1-4) solves a stated optimal-transport assignment, prototype updates (Eqs. 5-6) and losses (Eqs. 7-9) are explicit functions of features and prototypes, and inference (Eq. 10) is nearest-prototype matching. Accuracy numbers on FG3D and ModelNet40 are comparisons against independent baselines, not predictions derived from the model's own fitted values. No parameter is fitted to a subset and then reported as a prediction of a closely related quantity; no load-bearing claim is justified solely by a self-citation (the paper contains no such citation chain); the Sinkhorn/APDAGD solvers are external. Hyperparameter choices (K=20, alpha=0.2, eta0=0.999) are selected via ablation on the Airplane split; this is mild selection, not circular derivation. The skeptical concern that Eq. (6) averages all views of a class rather than only assigned views is an internal-consistency/correctness issue, not a circularity: even if the equation degenerates, the reported gains would be unexplained rather than tautological.

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

The central claim rests on four tuned hyperparameters, several standard algorithmic tools (Sinkhorn, EMA, contrastive loss), and two domain assumptions about clusterability and balance. No new physical or conceptual entities are introduced beyond prototypes, which are learned from data.

free parameters (4)
  • K (number of prototypes per class) = 20
    Selected via ablation on FG3D-Airplane (Table 4b); the final reported SOTA results use K=20.
  • tau (contrastive temperature) = 0.1
    Chosen by hand in Section 4.1; no ablation is shown for tau.
  • alpha (loss weight) = 0.2
    Selected via ablation on FG3D-Airplane (Table 4d); balances L_ce and L_pc.
  • eta_0 (initial momentum) = 0.999
    Selected via ablation (Table 4c); the momentum schedule eta_t = min(0.999, 1 - 1/(t+1)) for t > T is also hand-chosen and not derived.
assumptions (4)
  • standard math Sinkhorn-Knopp iteration converges to the solution of the entropy-regularized optimal transport problem (Eq. 4).
    Used to compute prototype assignment Z^c; this is a standard result from [50], not proven in the paper.
  • standard math Exponential moving average updates track cluster centroids in the embedding space (Eq. 5).
    The EMA recursion is assumed to converge to stable prototype locations; no convergence analysis is given.
  • domain assumption Fine-grained subcategories are approximately separable into K spherical clusters in the learned embedding space.
    Nearest-prototype inference and intra-class alignment assume each class can be represented by K centroids; the cluster structure is not validated.
  • domain assumption The balanced equipartition constraint (each prototype receives equal mass) is beneficial even for imbalanced subclasses.
    The authors impose balance to mitigate class imbalance, but this may distort natural cluster sizes and is not empirically isolated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Proto-FG3D: Prototype-based Interpretable Fine-Grained 3D Shape Classification." pith.science (2026). https://pith.science/paper/UHUXKIAM

@misc{pith2026250517666,
  author       = {Pith},
  title        = {Pith review of: Proto-FG3D: Prototype-based Interpretable Fine-Grained 3D Shape Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UHUXKIAM}},
  note         = {Machine review of arXiv:2505.17666}
}
read the original abstract

Deep learning-based multi-view coarse-grained 3D shape classification has achieved remarkable success over the past decade, leveraging the powerful feature learning capabilities of CNN-based and ViT-based backbones. However, as a challenging research area critical for detailed shape understanding, fine-grained 3D classification remains understudied due to the limited discriminative information captured during multi-view feature aggregation, particularly for subtle inter-class variations, class imbalance, and inherent interpretability limitations of parametric model. To address these problems, we propose the first prototype-based framework named Proto-FG3D for fine-grained 3D shape classification, achieving a paradigm shift from parametric softmax to non-parametric prototype learning. Firstly, Proto-FG3D establishes joint multi-view and multi-category representation learning via Prototype Association. Secondly, prototypes are refined via Online Clustering, improving both the robustness of multi-view feature allocation and inter-subclass balance. Finally, prototype-guided supervised learning is established to enhance fine-grained discrimination via prototype-view correlation analysis and enables ad-hoc interpretability through transparent case-based reasoning. Experiments on FG3D and ModelNet40 show Proto-FG3D surpasses state-of-the-art methods in accuracy, transparent predictions, and ad-hoc interpretability with visualizations, challenging conventional fine-grained 3D recognition approaches.

Figures

Figures reproduced from arXiv: 2505.17666 by the authors.

Figure 1
Figure 1. Multi-view 3D shape classification paradigms: (a) Parametric softmax can be interpreted as a learnable [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Architecture of prototype-based fine-grained 3D shape classification model. The visualization of initial [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 40 canonical work pages

  1. [1]

    Multi-view convolutional neural networks for 3d shape recognition

    Hang Su, Subhransu Maji, Evangelos Kalogerakis, and Erik Learned-Miller. Multi-view convolutional neural networks for 3d shape recognition. InProceedings of the IEEE international conference on computer vision, pages 945–953, 2015

  2. [2]

    Gvcnn: Group-view convolutional neural networks for 3d shape recognition

    Yifan Feng, Zizhao Zhang, Xibin Zhao, Rongrong Ji, and Yue Gao. Gvcnn: Group-view convolutional neural networks for 3d shape recognition. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 264–272, 2018

  3. [3]

    A deeper look at 3d shape classifiers

    Jong-Chyi Su, Matheus Gadelha, Rui Wang, and Subhransu Maji. A deeper look at 3d shape classifiers. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018

  4. [4]

    3d2seqviews: Aggregating sequential views for 3d global feature learning by cnn with hierarchical attention aggregation.IEEE Transactions on Image Processing, 28(8):3986–3999, 2019

    Zhizhong Han, Honglei Lu, Zhenbao Liu, Chi-Man V ong, Yu-Shen Liu, Matthias Zwicker, Junwei Han, and CL Philip Chen. 3d2seqviews: Aggregating sequential views for 3d global feature learning by cnn with hierarchical attention aggregation.IEEE Transactions on Image Processing, 28(8):3986–3999, 2019. 8 APREPRINT- SEPTEMBER13, 2025

  5. [5]

    End-to-end learning local multi-view descriptors for 3d point clouds

    Lei Li, Siyu Zhu, Hongbo Fu, Ping Tan, and Chiew-Lan Tai. End-to-end learning local multi-view descriptors for 3d point clouds. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1919–1928, 2020

  6. [6]

    Veram: View-enhanced recurrent attention model for 3d shape classification.IEEE transactions on visualization and computer graphics, 25(12):3244–3257, 2018

    Songle Chen, Lintao Zheng, Yan Zhang, Zhixin Sun, and Kai Xu. Veram: View-enhanced recurrent attention model for 3d shape classification.IEEE transactions on visualization and computer graphics, 25(12):3244–3257, 2018

  7. [7]

    Seqviews2seqlabels: Learning 3d global features via aggregating sequential views by rnn with attention.IEEE Transactions on Image Processing, 28(2):658–672, 2018

    Zhizhong Han, Mingyang Shang, Zhenbao Liu, Chi-Man V ong, Yu-Shen Liu, Matthias Zwicker, Junwei Han, and CL Philip Chen. Seqviews2seqlabels: Learning 3d global features via aggregating sequential views by rnn with attention.IEEE Transactions on Image Processing, 28(2):658–672, 2018

  8. [8]

    Mlvcnn: Multi-loop-view convolutional neural network for 3d shape retrieval

    Jianwen Jiang, Di Bao, Ziqiang Chen, Xibin Zhao, and Yue Gao. Mlvcnn: Multi-loop-view convolutional neural network for 3d shape retrieval. InProceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 8513–8520, 2019

Show all 52 references
  1. [9]

    View-gcn: View-based graph convolutional network for 3d shape analysis

    Xin Wei, Ruixuan Yu, and Jian Sun. View-gcn: View-based graph convolutional network for 3d shape analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1850–1859, 2020

  2. [10]

    Learning canonical view representation for 3d shape recognition with arbitrary views

    Xin Wei, Yifei Gong, Fudong Wang, Xing Sun, and Jian Sun. Learning canonical view representation for 3d shape recognition with arbitrary views. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 407–416, 2021

  3. [11]

    Mvt: Multi-view vision transformer for 3d object recognition

    Shuo Chen, Tan Yu, and Ping Li. Mvt: Multi-view vision transformer for 3d object recognition. InProceedings of the British Machine Vision Conference (BMVC), 2021

  4. [12]

    Vsformer: Mining correlations in flexible view set for multi-view 3d shape understanding.IEEE Transactions on Visualization and Computer Graphics, 2024

    Hongyu Sun, Yongcai Wang, Peng Wang, Haoran Deng, Xudong Cai, and Deying Li. Vsformer: Mining correlations in flexible view set for multi-view 3d shape understanding.IEEE Transactions on Visualization and Computer Graphics, 2024

  5. [13]

    Group multi-view transformer for 3d shape analysis with spatial encoding.IEEE Transactions on Multimedia, 2024

    Lixiang Xu, Qingzhe Cui, Richang Hong, Wei Xu, Enhong Chen, Xin Yuan, Chenglong Li, and Yuanyan Tang. Group multi-view transformer for 3d shape analysis with spatial encoding.IEEE Transactions on Multimedia, 2024

  6. [14]

    Deep inside convolutional networks: Visualising image classification models and saliency maps.arXiv preprint arXiv:1312.6034, 2013

    Karen Simonyan, Andrea Vedaldi, and Andrew Zisserman. Deep inside convolutional networks: Visualising image classification models and saliency maps.arXiv preprint arXiv:1312.6034, 2013

  7. [15]

    Synthesizing the preferred inputs for neurons in neural networks via deep generator networks.Advances in neural information processing systems, 29, 2016

    Anh Nguyen, Alexey Dosovitskiy, Jason Yosinski, Thomas Brox, and Jeff Clune. Synthesizing the preferred inputs for neurons in neural networks via deep generator networks.Advances in neural information processing systems, 29, 2016

  8. [16]

    Activation atlas.Distill, 4(3):e15, 2019

    Shan Carter, Zan Armstrong, Ludwig Schubert, Ian Johnson, and Chris Olah. Activation atlas.Distill, 4(3):e15, 2019

  9. [17]

    Hmtn: Hierarchical multi-scale transformer network for 3d shape recognition

    Yue Zhao, Weizhi Nie, Zan Gao, and An-an Liu. Hmtn: Hierarchical multi-scale transformer network for 3d shape recognition. InProceedings of the 30th ACM International Conference on Multimedia, pages 316–324, 2022

  10. [18]

    Mhsan: Multi-view hierarchical self-attention network for 3d shape recognition.Pattern Recognition, 150:110315, 2024

    Jiangzhong Cao, Lianggeng Yu, Bingo Wing-Kuen Ling, Zijie Yao, and Qingyun Dai. Mhsan: Multi-view hierarchical self-attention network for 3d shape recognition.Pattern Recognition, 150:110315, 2024

  11. [19]

    Fine-grained 3d shape classification with hierarchical part-view attention.IEEE Transactions on Image Processing, 30:1744–1758, 2021

    Xinhai Liu, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Fine-grained 3d shape classification with hierarchical part-view attention.IEEE Transactions on Image Processing, 30:1744–1758, 2021

  12. [20]

    Steps toward artificial intelligence.Proceedings of the IRE, 49(1):8–30, 1961

    Marvin Minsky. Steps toward artificial intelligence.Proceedings of the IRE, 49(1):8–30, 1961

  13. [21]

    Human problem solving.Upper Saddle River/Prentive Hall, 1972

    Allen Newell. Human problem solving.Upper Saddle River/Prentive Hall, 1972

  14. [22]

    Matching categorical object representations in inferior temporal cortex of man and monkey.Neuron, 60(6):1126–1141, 2008

    Nikolaus Kriegeskorte, Marieke Mur, Douglas A Ruff, Roozbeh Kiani, Jerzy Bodurka, Hossein Esteky, Keiji Tanaka, and Peter A Bandettini. Matching categorical object representations in inferior temporal cortex of man and monkey.Neuron, 60(6):1126–1141, 2008

  15. [23]

    Improving generalization via scalable neighborhood component analysis

    Zhirong Wu, Alexei A Efros, and Stella X Yu. Improving generalization via scalable neighborhood component analysis. InProceedings of the european conference on computer vision (ECCV), pages 685–701, 2018

  16. [24]

    Robust classification with convolutional prototype learning

    Hong-Ming Yang, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Robust classification with convolutional prototype learning. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3474–3482, 2018

  17. [25]

    Visual recognition with deep nearest centroids

    Wenguan Wang, Cheng Han, Tianfei Zhou, and Dongfang Liu. Visual recognition with deep nearest centroids. arXiv preprint arXiv:2209.07383, 2022

  18. [26]

    Prototypical networks for few-shot learning.Advances in neural information processing systems, 30, 2017

    Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning.Advances in neural information processing systems, 30, 2017. 9 APREPRINT- SEPTEMBER13, 2025

  19. [27]

    Hyperspherical prototype networks.Advances in neural information processing systems, 32, 2019

    Pascal Mettes, Elise Van der Pol, and Cees Snoek. Hyperspherical prototype networks.Advances in neural information processing systems, 32, 2019

  20. [28]

    Yonglong Tian, Yue Wang, Dilip Krishnan, Joshua B Tenenbaum, and Phillip Isola. Rethinking few-shot image classification: a good embedding is all you need? InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16, pages 2...

  21. [29]

    Prototypical priors: From improving classification to zero-shot learning.arXiv preprint arXiv:1512.01192, 2015

    Saumya Jetley, Bernardino Romera-Paredes, Sadeep Jayasumana, and Philip Torr. Prototypical priors: From improving classification to zero-shot learning.arXiv preprint arXiv:1512.01192, 2015

  22. [30]

    Attribute prototype network for zero-shot learning.Advances in Neural Information Processing Systems, 33:21969–21980, 2020

    Wenjia Xu, Yongqin Xian, Jiuniu Wang, Bernt Schiele, and Zeynep Akata. Attribute prototype network for zero-shot learning.Advances in Neural Information Processing Systems, 33:21969–21980, 2020

  23. [31]

    Panet: Few-shot image semantic segmentation with prototype alignment

    Kaixin Wang, Jun Hao Liew, Yingtian Zou, Daquan Zhou, and Jiashi Feng. Panet: Few-shot image semantic segmentation with prototype alignment. Inproceedings of the IEEE/CVF international conference on computer vision, pages 9197–9206, 2019

  24. [32]

    Adaptive prototype learning and allocation for few-shot segmentation

    Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive prototype learning and allocation for few-shot segmentation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8334–8343, 2021

  25. [33]

    Clustering based point cloud represen- tation learning for 3d analysis

    Tuo Feng, Wenguan Wang, Xiaohan Wang, Yi Yang, and Qinghua Zheng. Clustering based point cloud represen- tation learning for 3d analysis. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8283–8294, 2023

  26. [34]

    Prototype-based semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    Tianfei Zhou and Wenguan Wang. Prototype-based semantic segmentation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

  27. [35]

    Unsupervised feature learning via non-parametric instance discrimination

    Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 3733–3742, 2018

  28. [36]

    Prototypical contrastive learning of unsupervised representations.arXiv preprint arXiv:2005.04966, 2020

    Junnan Li, Pan Zhou, Caiming Xiong, and Steven CH Hoi. Prototypical contrastive learning of unsupervised representations.arXiv preprint arXiv:2005.04966, 2020

  29. [37]

    Augmentation invariant and instance spreading feature for softmax embedding.IEEE transactions on pattern analysis and machine intelligence, 44(2):924–939, 2020

    Mang Ye, Jianbing Shen, Xu Zhang, Pong C Yuen, and Shih-Fu Chang. Augmentation invariant and instance spreading feature for softmax embedding.IEEE transactions on pattern analysis and machine intelligence, 44(2):924–939, 2020

  30. [38]

    Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions

    Oscar Li, Hao Liu, Chaofan Chen, and Cynthia Rudin. Deep learning for case-based reasoning through prototypes: A neural network that explains its predictions. InProceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018

  31. [39]

    Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds

    Tuo Feng, Ruijie Quan, Xiaohan Wang, Wenguan Wang, and Yi Yang. Interpretable3d: An ad-hoc interpretable classifier for 3d point clouds. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 1761–1769, 2024

  32. [40]

    Learning deep features for discriminative localization

    Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features for discriminative localization. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2921–2929, 2016

  33. [41]

    Grad-cam: Why did you say that?arXiv preprint arXiv:1611.07450, 2016

    Ramprasaath R Selvaraju, Abhishek Das, Ramakrishna Vedantam, Michael Cogswell, Devi Parikh, and Dhruv Batra. Grad-cam: Why did you say that?arXiv preprint arXiv:1611.07450, 2016

  34. [42]

    Transformer interpretability beyond attention visualization

    Hila Chefer, Shir Gur, and Lior Wolf. Transformer interpretability beyond attention visualization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 782–791, 2021

  35. [43]

    Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study.PLoS medicine, 15(11):e1002683, 2018

    John R Zech, Marcus A Badgeley, Manway Liu, Anthony B Costa, Joseph J Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: a cross-sectional study.PLoS medicine, 15(11):e1002683, 2018

  36. [44]

    Dan: Deep-attention network for 3d shape recognition.IEEE Transactions on Image Processing, 30:4371–4383, 2021

    Weizhi Nie, Yue Zhao, Dan Song, and Yue Gao. Dan: Deep-attention network for 3d shape recognition.IEEE Transactions on Image Processing, 30:4371–4383, 2021

  37. [45]

    Multi-view classification using hybrid fusion and mutual distillation

    Samuel Black and Richard Souvenir. Multi-view classification using hybrid fusion and mutual distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 270–280, 2024

  38. [46]

    This looks like that: deep learning for interpretable image recognition.Advances in neural information processing systems, 32, 2019

    Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan K Su. This looks like that: deep learning for interpretable image recognition.Advances in neural information processing systems, 32, 2019

  39. [47]

    Protopformer: concentrating on prototypical parts in vision transformers for interpretable image recognition

    Mengqi Xue, Qihan Huang, Haofei Zhang, Jingwen Hu, Jie Song, Mingli Song, and Canghong Jin. Protopformer: concentrating on prototypical parts in vision transformers for interpretable image recognition. InProceedings of the Thirty-Third International Joint Conference on Artific...

  40. [48]

    Protopshare: Prototypical parts sharing for similarity discovery in interpretable image classification

    Dawid Rymarczyk, Łukasz Struski, Jacek Tabor, and Bartosz Zieli´nski. Protopshare: Prototypical parts sharing for similarity discovery in interpretable image classification. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 1420–1430, 2021

  41. [49]

    Neural prototype trees for interpretable fine-grained image recognition

    Meike Nauta, Ron Van Bree, and Christin Seifert. Neural prototype trees for interpretable fine-grained image recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14933–14943, 2021

  42. [50]

    Lightspeed computation of optimal transport.Advances in neural information processing systems, 26:2292–2300, 2013

    Cuturi M Sinkhorn Distances. Lightspeed computation of optimal transport.Advances in neural information processing systems, 26:2292–2300, 2013

  43. [51]

    On efficient optimal transport: An analysis of greedy and accelerated mirror descent algorithms

    Tianyi Lin, Nhat Ho, and Michael Jordan. On efficient optimal transport: An analysis of greedy and accelerated mirror descent algorithms. InInternational Conference on Machine Learning, pages 3982–3991. PMLR, 2019

  44. [52]

    3d shapenets: A deep representation for volumetric shapes

    Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaoou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1912–1920, 2015. 11

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.