Pith. sign in

REVIEW 4 major objections 5 minor 57 references

Adversarial Reconstruction Feedback for Robust Fine-grained Generalization

T0 review · 4 major / 5 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 This paper claims that fine-grained image retrieval generalizes to unseen subcategories when visual discrepancies are reconstructed from appearance cues alone, and that its AdvRF framework achieves this through an adversarial feedback…

desk verdict A plausible adversarial reconstruction loop for fine-grained retrieval with consistent ablations, but the category-agnostic mechanism is asserted rather than measured. read the letter →

arxiv 2507.21742 v1 pith:NN3UMG3B submitted 2025-07-29 cs.CV

Shijie Wang , Jian Shi , Haojie Li This is my paper
classification cs.CV
keywords fine-grainedimageretrievalcategory-agnosticrepresentationadversarialfeedbackreconstructionmodelknowledgedistillationdiscrepancylocalizationunseencategoriesRecall@Kevaluation
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

This paper tries to show that fine-grained image retrieval fails on unseen categories because training on fixed classes leaks category semantics into the embeddings. It proposes AdvRF, an adversarial loop between a retrieval model and a reconstruction model: the retrieval model marks where the visual discrepancies are, and the reconstruction model re-represents exactly those regions from appearance alone. The two models take turns: the reconstructor exposes residual discrepancies the retriever missed, and the retriever's refined localization forces harder reconstruction. The purified category-agnostic discrepancy representation is then distilled back into the retriever, so at test time only the light retrieval model runs. If correct, the framework would make retrieval embeddings describe visual appearance rather than memorized class labels, which matters because real-world queries often involve subcategories the model has never seen.

What carries the argument

The central object is the masked discrepancy representation $C_A = F_I \odot M$, built from the reconstruction encoder's feature map $F_I$ and a pattern map $M = \sigma(E(T)(F))$ derived from the retrieval feature $F$. The mask separates appearance-level discrepancies from the rest of the image; the adversarial feedback alternates a reconstruction loss $L^G_{\mathrm{RecF}}$ from the reconstruction model's residual errors with a retrieval loss $L^{\mathrm{RetF}}_R$ that maximizes the reconstruction error caused by the current localization, finally transferring the purified representation into the retrieval embedding through $L_P = \|E_R - g(C_A)\|$. All of the reported generalization behaviour is attributed to this alternating cycle.

What would settle it

Set the pattern map to a constant (uniform) value and retrain with the same losses; if the masked version does not beat the unmasked version on held-out subcategories, the category-agnostic target is not what carries the generalization. Alternatively, shuffle the training-set category labels and measure whether nearest-neighbour rankings under the trained embeddings change; large changes would indicate that category semantics persist in the embedding.

Watch

Extended reading notes

Core claim

The central claim is that fine-grained retrieval embeddings trained with classification supervision are contaminated by category-specific semantics, and that replacing the discrepancy target with a category-agnostic reconstruction target removes that contamination. AdvRF operationalizes this by splitting the input representation into discrepancy and non-discrepancy parts according to a mask produced by the retrieval model, reconstructing each part separately, and using the reconstruction failures as feedback to sharpen localization. The purified discrepancy representation is distilled into the retrieval embedding, so inference needs only the retrieval backbone. The paper reports Recall@1 of 76.6% on CUB-200-2011, 94.9% on Stanford Cars, 88.0% on FGVC Aircraft, and 84.2% on Stanford Online Products, each above the best prior numbers it compares against.

Load-bearing premise

The load-bearing premise is that the reconstruction encoder's masked features are genuinely category-agnostic, but the paper never directly measures that property; if the mask still selects category-level regions, the gains could come from the extra losses rather than from removal of category semantics.

Editorial extensions

If this is right

  • At deployment the framework needs only the retrieval model, so the reported recall comes at the cost of a single backbone forward pass; the reconstruction and distillation machinery are train-time only.
  • Because the target is appearance-based rather than class-logit-based, the retrieval embedding can in principle be transferred to a new taxonomy without retraining on its classes, which is exactly the unseen-category setting the benchmarks simulate.
  • The ablation showing that both discrepancy and non-discrepancy reconstruction losses are needed implies that the residual-error signal, not just the reconstruction objective itself, drives the localization improvement.
  • If the adversarial cycle is what matters, performance should keep improving when the reconstructor is made more sensitive to small residuals, since stronger residual feedback forces sharper masks.

Reading between the lines

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

  • Because the reconstruction target needs no class annotations beyond the initial classification loss, the same feedback loop should transfer to few-shot or partially labelled retrieval settings, where class supervision is scarce.
  • A direct probe of the category-agnostic claim would permute the training labels before the final distillation and check whether nearest-neighbour rankings change; stable rankings would confirm appearance-based encoding.
  • The adversarial feedback principle could be applied beyond images, for example to fine-grained retrieval of audio or text, wherever a reconstruction model can expose residuals that a trained retriever misses.
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 / 5 minor

Summary. The paper proposes AdvRF, an adversarial reconstruction feedback framework for fine-grained image retrieval. The method alternates between updating a retrieval model, which localizes visual discrepancies via a learned pattern map, and a reconstruction model, which encodes and reconstructs discrepancy and non-discrepancy regions. A knowledge-distillation loss transfers the reconstruction model's pooled discrepancy features into the retrieval embedding. The authors evaluate on CUB-200-2011, Stanford Cars, FGVC Aircraft, and Stanford Online Products, reporting state-of-the-art Recall@k results and ablations showing that each loss component contributes to the gains.

Significance. If the central mechanism is sound, the paper offers a novel way to decouple discrepancy modeling from predefined category supervision, which could benefit open-set and zero-shot retrieval. The paper presents a clear formulation, a logical ablation across loss components, and evaluations on multiple benchmarks. However, the core assumption that the reconstruction features are category-agnostic is not directly verified, and the reported gains over strong baselines are modest without error bars. The paper would be substantially strengthened by direct quantitative measurements of category information in the learned embeddings and by multi-run statistics.

major comments (4)
  1. [Sec. 3.2, Eq. (5)] The loss LP = ||ER - g(CA)|| is presented as ensuring that ER 'exclusively capture[s] category-agnostic visual discrepancies.' However, CA = FI ⊙ M, and M is produced by σ(E(T)(F)) (Eq. 2), where F is the category-supervised retrieval feature. Thus the distillation target itself depends on the category-supervised network, and no term in Eqs. (5), (14), or (15) penalizes category information in FI or M. Please provide a quantitative test of category-agnosticism, e.g., linear probing accuracy of ER and FI with respect to the training categories, or an ablation that uses a mask from an unsupervised source, to support the claim that category semantics are removed.
  2. [Table 4] The IDML row reports R@100 = 54.8 and R@1000 = 51.3, which are far below its own R@1 = 81.5 and R@10 = 92.3. This is internally inconsistent and likely a typo. Please correct the values and re-verify all numbers in the table, because this inconsistency undermines the reliability of the comparison with the proposed method.
  3. [Tables 1, 3, and 4] All results are reported as single numbers without variance or multiple seeds. Given that the final improvements over strong baselines are modest (e.g., +1.8 on CUB R@1, +0.5 on Cars, +1.8 on SOP R@1), it is essential to report mean ± standard deviation over at least three runs, or to provide significance tests, so that the reader can judge whether the gains are meaningful.
  4. [Sec. 4.4, Fig. 5] The 'Analysis of category-agnostic discrepancies' is qualitative and does not directly measure category-agnosticism. The figure shows that category-agnostic embeddings retrieve same-subcategory samples for novel classes, which is a retrieval outcome, not evidence that category semantics are absent. A linear probe or mutual information estimate on the learned ER (and on FI) for the training categories would directly test the central claim. Without such evidence, the performance gains could plausibly be attributed to the reconstruction and consistency losses acting as regularizers.
minor comments (5)
  1. [Abstract and Sec. 1] The phrase 'significantly outperforming existing state-of-the-art methods' is stronger than the data support, since gains over the best baselines are around 1–2 points in Recall@1. Please temper the claim or provide statistical evidence.
  2. [Sec. 3.2, Eq. (3)] The EMA update notation is confusing. Please clarify whether E(t)(T) on the left-hand side is the updated running average, and define the role of δ relative to conventional momentum formulations.
  3. [Sec. 4.1, CUB split] The sentence 'and the consists of (5,924 images) for testing' contains a typo. Please also confirm that the standard 100/100 train/test split for CUB is used.
  4. [Sec. 3.1, decoder description] The decoder is described as a 'U-Net architecture,' but no skip connections are mentioned. Please clarify whether the decoder actually receives encoder features at multiple resolutions, or whether it is a plain convolutional autoencoder.
  5. [Sec. 4.2, Table 5] The text states that 'as the ratio increases, retrieval performance declines,' but the table shows δ = 0.2 (76.6%) outperforms δ = 0.1 (75.4%), so the trend is not monotonic. Please revise the description to reflect the actual pattern and explain the non-monotonicity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark results are externally grounded; the category-agnostic mechanism is asserted rather than measured, but it does not reduce the reported gains to the paper's inputs.

full rationale

The paper's central quantitative claims are evaluated on unseen test categories (CUB, Cars, Aircraft, SOP) outside the training loop, so the reported Recall@K improvements are not forced by construction from the loss equations. The closest concern is the 'category-agnostic' mechanism: Sec. 3.2 assumes the reconstruction encoder yields 'the category-agnostic representation F_I', defines C_A = F_I ⊙ M in Eq. 4, and then Eq. 5 trains E_R toward g(C_A), with M itself produced from the category-supervised features F via Eq. 2. This makes the distillation target self-referential, and the paper never directly measures category information in F_I or E_R; it concedes in Sec. 4.4 that it uses 'an indirect method to interpret category-agnostic discrepancies.' That is a missing-support or validity weakness, and it weakens the mechanistic explanation, but it is not a circular reduction of the experimental outcomes: the gains come from an empirical comparison with external baselines, and the ablation in Tab. 2 shows a non-trivial difference between feedback types. There is also no load-bearing self-citation or imported uniqueness theorem; the reconstruction-is-category-agnostic premise rests on external citations [9, 12]. Overall, no specific equation or fitted parameter is equivalent to the prediction by construction, so the circularity score is 0.

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

The framework's central mechanism rests on four hyperparameters tuned on CUB and several unproved modeling assumptions about reconstruction features being category-agnostic. These assumptions and tuned values are the price of the claimed generalization.

free parameters (4)
  • alpha = 0.7
    Loss weight for reconstruction feedback in Eq 14, tuned on CUB-200-2011 (Fig 3).
  • beta = 0.5
    Loss weight for LP in Eq 15, tuned on CUB-200-2011 (Fig 3).
  • gamma = 0.6
    Loss weight for retrieval feedback in Eq 15, tuned on CUB-200-2011 (Fig 3).
  • delta = 0.2
    EMA update ratio for the mean generator in Eq 3, chosen from a sweep in Table 5 (best R@1=76.6).
assumptions (4)
  • domain assumption Reconstruction models inherently learn category-agnostic representations based on pixel-level fidelity.
    Invoked in Sec 1 and Sec 3.2 as the motivation for using GE's features as category-agnostic discrepancy representations; no direct measurement of category-agnosticism is provided.
  • domain assumption The adversarial alternating optimization (RecU/RetU) converges to a stable, non-degenerate equilibrium where the mask M localizes truthful discrepancies.
    Sec 3.4 relies on this without convergence analysis; degenerate masks (e.g., all-ones or random noise) are not theoretically ruled out.
  • domain assumption The retrieval features F contain sufficient spatial information for the lightweight generator T to produce an accurate discrepancy mask.
    Eqs 1 and 2 assume that a 1x1 convolution on F can localize meaningful discrepancies; no formal analysis of mask quality is given beyond visual examples.
  • domain assumption ImageNet-pretrained backbones (ResNet-50, ResNet-34) provide a good starting point for both retrieval and reconstruction.
    Sec 4.1 uses ImageNet weights without measuring their contribution; the reconstruction encoder inherits ImageNet semantics, which may already embed category-relevant information.
invented entities (2)
  • Pattern map M
    purpose: Spatial mask that localizes visual discrepancies and gates the decomposition of reconstruction features into CA and CR.
    M is produced by a learned generator T on retrieval features; its quality is shown only qualitatively in Fig 4 and via a comparison in Table 6. No independent validation outside the framework.
  • Category-agnostic discrepancy representation CA
    purpose: The masked reconstruction feature intended to encode visual differences without category semantics; used as a distillation target for the retrieval embedding.
    The category-agnostic property is asserted from reconstruction literature, not demonstrated on this framework; CA is an internal latent variable whose agnosticism is not directly measured.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adversarial Reconstruction Feedback for Robust Fine-grained Generalization." pith.science (2026). https://pith.science/paper/NN3UMG3B

@misc{pith2026250721742,
  author       = {Pith},
  title        = {Pith review of: Adversarial Reconstruction Feedback for Robust Fine-grained Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NN3UMG3B}},
  note         = {Machine review of arXiv:2507.21742}
}
read the original abstract

Existing fine-grained image retrieval (FGIR) methods predominantly rely on supervision from predefined categories to learn discriminative representations for retrieving fine-grained objects. However, they inadvertently introduce category-specific semantics into the retrieval representation, creating semantic dependencies on predefined classes that critically hinder generalization to unseen categories. To tackle this, we propose AdvRF, a novel adversarial reconstruction feedback framework aimed at learning category-agnostic discrepancy representations. Specifically, AdvRF reformulates FGIR as a visual discrepancy reconstruction task via synergizing category-aware discrepancy localization from retrieval models with category-agnostic feature learning from reconstruction models. The reconstruction model exposes residual discrepancies overlooked by the retrieval model, forcing it to improve localization accuracy, while the refined signals from the retrieval model guide the reconstruction model to improve its reconstruction ability. Consequently, the retrieval model localizes visual differences, while the reconstruction model encodes these differences into category-agnostic representations. This representation is then transferred to the retrieval model through knowledge distillation for efficient deployment. Quantitative and qualitative evaluations demonstrate that our AdvRF achieves impressive performance on both widely-used fine-grained and coarse-grained datasets.

Figures

Figures reproduced from arXiv: 2507.21742 by the authors.

Figure 1
Figure 1. Motivation of the proposed AdvRF. (a) Classification [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Detailed illustration of adversarial reconstruction feedback. See §3 for more details. thermore, these discrepancies are back into the retrieval model via knowledge distillation, significantly boosting computational efficiency during test. 3.1. Network Architecture Fine-grained Retrieval Model R. It is designed to ex￾tract robust object representations and generate the final retrieval embeddings. Given an input imag… view at source ↗
Figure 5
Figure 5. Evaluating the efficacy of category-related embeddings [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualizations of pattern maps using different feedback: [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 54 canonical work pages

  1. [1]

    Ak, Ashraf A

    Kenan E. Ak, Ashraf A. Kassim, Joo-Hwee Lim, and Jo Yew Tham. Learning attribute representations with localiza- tion for flexible fashion search. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018 , pages 7708–

    work page 2018

  2. [2]

    A unifying mutual information view of metric learn- ing: Cross-entropy vs

    Malik Boudiaf, J ´erˆome Rony, Imtiaz Masud Ziko, Eric Granger, Marco Pedersoli, Pablo Piantanida, and Ismail Ben Ayed. A unifying mutual information view of metric learn- ing: Cross-entropy vs. pairwise losses. In ECCV, pages 548–

    work page

  3. [3]

    Belongie, and Pietro Perona

    Steve Branson, Grant Van Horn, Serge J. Belongie, and Pietro Perona. Bird species categorization using pose nor- malized deep convolutional nets. CoRR, abs/1406.2952,

    arXiv

  4. [4]

    Robust and accurate object detection via adversarial learning

    Xiangning Chen, Cihang Xie, Mingxing Tan, Li Zhang, Cho- Jui Hsieh, and Boqing Gong. Robust and accurate object detection via adversarial learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, vir- tual, June 19-25, 2021 , pages 16622–16631. Computer Vi- sion Foundation / IEEE, 2021. 2

    work page 2021

  5. [5]

    Elgammal

    Mohamed Elhoseiny, Yizhe Zhu, Han Zhang, and Ahmed M. Elgammal. Link the head to the ”beak”: Zero shot learn- ing from noisy text description at part precision. In 2017 IEEE Conference on Computer Vision and Pattern Recog- nition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017 , pages 6288–6297. IEEE Computer Society, 2017. 1

    work page 2017

  6. [6]

    Masked auto-encoders meet generative adversarial networks and beyond

    Zhengcong Fei, Mingyuan Fan, Li Zhu, Junshi Huang, Xi- aoming Wei, and Xiaolin Wei. Masked auto-encoders meet generative adversarial networks and beyond. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023 , pages 24449–24459. IEEE, 2023. 2

    work page 2023

  7. [7]

    Lempitsky

    Yaroslav Ganin and Victor S. Lempitsky. Unsupervised do- main adaptation by backpropagation. In Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, pages 1180–1189. JMLR.org, 2015. 2

    work page 2015

  8. [8]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. Com- mun. ACM, 63(11):139–144, 2020. 2

    work page 2020

Show all 57 references
  1. [9]

    Image-based 3d object reconstruction: State-of-the-art and trends in the deep learning era

    Xian-Feng Han, Hamid Laga, and Mohammed Bennamoun. Image-based 3d object reconstruction: State-of-the-art and trends in the deep learning era. IEEE Trans. Pattern Anal. Mach. Intell., 43(5):1578–1604, 2021. 2

    2021

  2. [10]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016 , pages 770– 778, 2016. 5, 6, 8

    2016

  3. [11]

    Duplex generative adversarial network for unsupervised do- main adaptation

    Lanqing Hu, Meina Kan, Shiguang Shan, and Xilin Chen. Duplex generative adversarial network for unsupervised do- main adaptation. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1498–1507. Computer...

    2018

  4. [12]

    Few-view object reconstruction with unknown cate- gories and camera poses

    Hanwen Jiang, Zhenyu Jiang, Kristen Grauman, and Yuke Zhu. Few-view object reconstruction with unknown cate- gories and camera poses. In International Conference on 3D Vision, 3DV 2024, Davos, Switzerland, March 18-21, 2024 , pages 31–41. IEEE, 2024. 2

    2024

  5. [13]

    Contrastive bayesian analysis for deep metric learning

    Shichao Kan, Zhiquan He, Yigang Cen, Yang Li, Vladimir Mladenovic, and Zhihai He. Contrastive bayesian analysis for deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 45(6):7220–7238, 2023. 6

    2023

  6. [14]

    Embedding transfer with label relaxation for improved met- ric learning

    Sungyeon Kim, Dongwon Kim, Minsu Cho, and Suha Kwak. Embedding transfer with label relaxation for improved met- ric learning. In CVPR, pages 3967–3976. Computer Vision Foundation / IEEE, 2021. 2, 7

    2021

  7. [15]

    Learning with memory-based virtual classes for deep metric learning

    ByungSoo Ko, Geonmo Gu, Han-Gyu Kim, and ByungSoo Ko. Learning with memory-based virtual classes for deep metric learning. In ICCV, pages 11772–11781. IEEE, 2021. 2

    2021

  8. [16]

    Opengan: Open-set recog- nition via open data generation

    Shu Kong and Deva Ramanan. Opengan: Open-set recog- nition via open data generation. In 2021 IEEE/CVF Inter- national Conference on Computer Vision, ICCV 2021, Mon- treal, QC, Canada, October 10-17, 2021 , pages 793–802. IEEE, 2021. 2

    2021

  9. [17]

    3d object representations for fine-grained categorization

    Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013, pages 554–561, 2013. 5

    2013

  10. [18]

    Correlation-guided semantic consistency network for visible-infrared person re-identification

    Haojie Li, Mingxuan Li, Qijie Peng, Shijie Wang, Hong Yu, and Zhihui Wang. Correlation-guided semantic consistency network for visible-infrared person re-identification. IEEE Trans. Circuits Syst. Video Technol., 34(6):4503–4515, 2024. 1

    2024

  11. [19]

    Hypergraph-induced semantic tuplet loss for deep metric learning

    Jongin Lim, Sangdoo Yun, Seulki Park, and Jin Young Choi. Hypergraph-induced semantic tuplet loss for deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 212–222. IEEE, 2022. 1, 6, 7

    2022

  12. [20]

    DAS: densely- anchored sampling for deep metric learning

    Lizhao Liu, Shangxin Huang, Zhuangwei Zhuang, Ran Yang, Mingkui Tan, and Yaowei Wang. DAS: densely- anchored sampling for deep metric learning. InECCV, pages 399–417. Springer, 2022. 1, 6, 7

    2022

  13. [21]

    Deepfashion: Powering robust clothes recognition and retrieval with rich annotations

    Ziwei Liu, Ping Luo, Shi Qiu, Xiaogang Wang, and Xiaoou Tang. Deepfashion: Powering robust clothes recognition and retrieval with rich annotations. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV , USA, June 27-30, 2016, pages 1096–...

    2016

  14. [22]

    Blaschko, and Andrea Vedaldi

    Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew B. Blaschko, and Andrea Vedaldi. Fine-grained visual classi- fication of aircraft. CoRR, abs/1306.5151, 2013. 5

    2013 arXiv

  15. [23]

    Keypoint-aligned embeddings for image retrieval and re-identification

    Olga Moskvyak, Fr ´ed´eric Maire, Feras Dayoub, and Mahsa Baktashmotlagh. Keypoint-aligned embeddings for image retrieval and re-identification. In WACV, pages 676–685. IEEE, 2021. 1, 2

    2021

  16. [24]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Inter- vention - MICCAI 2015 - 18th International Conference Mu- nich, Germany, October 5 - 9, 2015, Proceeding...

    2015

  17. [25]

    Simultaneous similarity-based self-distillation for deep metric learning

    Karsten Roth, Timo Milbich, Bj ¨orn Ommer, Joseph Paul Co- hen, and Marzyeh Ghassemi. Simultaneous similarity-based self-distillation for deep metric learning. In ICML, pages 9095–9106. PMLR, 2021. 2

    2021

  18. [26]

    Non- isotropy regularization for proxy-based deep metric learning

    Karsten Roth, Oriol Vinyals, and Zeynep Akata. Non- isotropy regularization for proxy-based deep metric learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18- 24, 2022, pages 7410–7420. IEEE, 2022. 1, 2, 6

    2022

  19. [27]

    Facenet: A unified embedding for face recognition and clus- tering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clus- tering. In CVPR, pages 815–823. IEEE Computer Society,

  20. [28]

    Learning intra-batch connections for deep metric learning

    Jenny Seidenschwarz. Learning intra-batch connections for deep metric learning. In ICML, pages 9410–9421. PMLR,

  21. [29]

    Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba- tra

    Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Ba- tra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336– 359, 2020. 7

    2020

  22. [30]

    Deep metric learning via lifted structured feature embedding

    Hyun Oh Song, Yu Xiang, Stefanie Jegelka, and Silvio Savarese. Deep metric learning via lifted structured feature embedding. In CVPR, pages 4004–4012. IEEE Computer Society, 2016. 6

    2016

  23. [31]

    Taylor, and Graham

    Eu Wern Teh, Terrance DeVries, Graham W. Taylor, and Graham. Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis. In ECCV, pages 448–

  24. [32]

    Introspective deep metric learning

    Chengkun Wang, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Introspective deep metric learning. IEEE Trans. Pattern Anal. Mach. Intell., 46(4):1964–1980, 2024. 6, 7

    1964

  25. [33]

    Category-specific nuance exploration network for fine- grained object retrieval

    Shijie Wang, Zhihui Wang, Haojie Li, and Wanli Ouyang. Category-specific nuance exploration network for fine- grained object retrieval. In Thirty-Sixth AAAI Conference on Artificial Intelligence, AAAI 2022, Thirty-Fourth Conference on Innovative Applications of Artificial Inte...

    2022

  26. [34]

    Open-set fine-grained retrieval via prompting vision-language evaluator

    Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Open-set fine-grained retrieval via prompting vision-language evaluator. In IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023 , pa...

    2023

  27. [35]

    Learning to parameterize visual at- tributes for open-set fine-grained retrieval

    Shijie Wang, Jianlong Chang, Haojie Li, Zhihui Wang, Wanli Ouyang, and Qi Tian. Learning to parameterize visual at- tributes for open-set fine-grained retrieval. In Advances in Neural Information Processing Systems 36: Annual Con- ference on Neural Information Processing Syste...

    2023

  28. [36]

    Fine-grained retrieval prompt tuning

    Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Fine-grained retrieval prompt tuning. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applica- tions of Artificial Intelligence, IAAI ...

    2023

  29. [37]

    Semantic-guided information align- ment network for fine-grained image recognition

    Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Semantic-guided information align- ment network for fine-grained image recognition. IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1, 2023. 1

    2023

  30. [38]

    Content-aware rectified activation for zero-shot fine-grained image retrieval

    Shijie Wang, Jianlong Chang, Zhihui Wang, Haojie Li, Wanli Ouyang, and Qi Tian. Content-aware rectified activation for zero-shot fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell., 46(6):4366–4380, 2024. 2, 6, 7

    2024

  31. [39]

    Accurate fine-grained object recogni- tion with structure-driven relation graph networks

    Shijie Wang, Zhihui Wang, Haojie Li, Jianlong Chang, Wanli Ouyang, and Qi Tian. Accurate fine-grained object recogni- tion with structure-driven relation graph networks. Int. J. Comput. Vis., 132(1):137–160, 2024. 1

    2024

  32. [40]

    Xun Wang, Xintong Han, Weilin Huang, Dengke Dong, and Matthew R. Scott. Multi-similarity loss with general pair weighting for deep metric learning. In CVPR, pages 5022–

  33. [41]

    Graph-propagation based correlation learning for weakly supervised fine-grained image classification

    Zhuhui Wang, Shijie Wang, Haojie Li, Zhi Dou, and Jianjun Li. Graph-propagation based correlation learning for weakly supervised fine-grained image classification. In The Thirty- Fourth AAAI Conference on Artificial Intelligence, AAAI 2020, The Thirty-Second Innovative Applica...

    2020

  34. [42]

    Weakly supervised fine-grained image classification via guassian mixture model oriented discrim- inative learning

    Zhihui Wang, Shijie Wang, Shuhui Yang, Haojie Li, Jianjun Li, and Zezhou Li. Weakly supervised fine-grained image classification via guassian mixture model oriented discrim- inative learning. In 2020 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, CVPR 2020, ...

    2020

  35. [43]

    Bppattack: Stealthy and efficient trojan attacks against deep neural net- works via image quantization and contrastive adversarial learning

    Zhenting Wang, Juan Zhai, and Shiqing Ma. Bppattack: Stealthy and efficient trojan attacks against deep neural net- works via image quantization and contrastive adversarial learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA,...

    2022

  36. [44]

    Selective convolutional descriptor aggregation for fine-grained image retrieval

    Xiu-Shen Wei, Jian-Hao Luo, Jianxin Wu, and Zhi-Hua Zhou. Selective convolutional descriptor aggregation for fine-grained image retrieval. IEEE Trans. Image Process. , 26(6):2868–2881, 2017. 2, 6

    2017

  37. [45]

    Attribute-aware deep hashing with self- consistency for large-scale fine-grained image retrieval

    Xiu-Shen Wei, Yang Shen, Xuhao Sun, Peng Wang, and Yuxin Peng. Attribute-aware deep hashing with self- consistency for large-scale fine-grained image retrieval. IEEE Trans. Pattern Anal. Mach. Intell. , 45(11):13904– 13920, 2023. 1

    2023

  38. [46]

    Bailin Yang, Haoqiang Sun, Frederick W. B. Li, Zheng Chen, Jianlu Cai, and Chao Song. HSE: hybrid species embedding for deep metric learning. In IEEE/CVF International Con- ference on Computer Vision, ICCV 2023, Paris, France, Oc- tober 1-6, 2023, pages 11013–11023. IEEE, 2023. 6, 7

    2023

  39. [47]

    Hard decorrelated centralized loss for fine-grained image retrieval

    Xianxian Zeng, Shun Liu, Xiaodong Wang, Yun Zhang, Kairui Chen, and Dong Li. Hard decorrelated centralized loss for fine-grained image retrieval. Neurocomputing, 453: 26–37, 2021. 6

    2021

  40. [48]

    Classification is a strong baseline for deep metric learning

    Andrew Zhai and Hao-Yu Wu. Classification is a strong baseline for deep metric learning. InBMVC, page 91. BMV A Press, 2019. 7

    2019

  41. [49]

    Deep compositional metric learning

    Wenzhao Zheng, Chengkun Wang, Jiwen Lu, and Jie Zhou. Deep compositional metric learning. In CVPR, pages 9320–

  42. [50]

    Deep relational metric learning

    Wenzhao Zheng, Borui Zhang, Jiwen Lu, and Jie Zhou. Deep relational metric learning. In ICCV, pages 12045–12054. IEEE, 2021. 2, 7

    2021

  43. [51]

    Centralized ranking loss with weakly supervised localization for fine-grained object retrieval

    Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Yongjian Wu, Feiyue Huang, and Yanhua Yang. Centralized ranking loss with weakly supervised localization for fine-grained object retrieval. In IJCAI, pages 1226–1233. ijcai.org, 2018. 2, 6

    2018

  44. [52]

    Towards optimal fine grained retrieval via decorrelated centralized loss with normalize-scale layer

    Xiawu Zheng, Rongrong Ji, Xiaoshuai Sun, Baochang Zhang, Yongjian Wu, and Feiyue Huang. Towards optimal fine grained retrieval via decorrelated centralized loss with normalize-scale layer. In AAAI, pages 9291–9298. AAAI Press, 2019. 6

    2019

  45. [53]

    Joint discriminative and generative learning for person re-identification

    Zhedong Zheng, Xiaodong Yang, Zhiding Yu, Liang Zheng, Yi Yang, and Jan Kautz. Joint discriminative and generative learning for person re-identification. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 2138...

    2019

  46. [464]

    Springer, 2020. 6, 7

    2020

  47. [5030]

    Computer Vision Foundation / IEEE, 2019. 7

    2019

  48. [7717]

    Computer Vision Foundation / IEEE Computer Soci- ety, 2018. 1

    2018

  49. [9329]

    Computer Vision Foundation / IEEE, 2021. 2, 7

    2021

Pith tools

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