Pith. sign in

REVIEW 4 major objections 6 minor 64 references

H3Former claims that a dynamic weighted hypergraph grouping image tokens into semantic regions, supervised by a hierarchical contrastive loss in hyperbolic space, sets new highs on four fine-grained visual classification benchmarks.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 22:26 UTC pith:XOCT376N

load-bearing objection The method is a plausible new combination, but the SOTA claim is undercut by protocol mismatch and an ablation loss defined only in the table, so it needs a major revision before the results can be taken at face value. the 4 major comments →

arxiv 2511.10260 v2 pith:XOCT376N submitted 2025-11-13 cs.CV cs.AI

H3Former: Hypergraph-based Semantic-Aware Aggregation via Hyperbolic Hierarchical Contrastive Loss for Fine-Grained Visual Classification

Yongji Zhang , Siqi Li , Kuiyang Huang , Yue Gao , Yu Jiang This is my paper
classification cs.CV cs.AI
keywords fine-grained visual classificationhypergraph neural networkvision transformersemantic-aware aggregationhyperbolic embeddinghierarchical contrastive learningtoken-to-regionLorentz model
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that fine-grained visual classification improves when a model groups image tokens into semantic regions rather than merely selecting informative tokens or proposing arbitrary regions. Its framework, H3Former, builds a dynamic weighted hypergraph over token features, aggregates tokens into region-level representations via hypergraph convolution, and supervises the resulting region hierarchy with a hyperbolic contrastive loss. The reported numbers—92.7% on CUB-200-2011, 91.6% on NA-Birds, 95.8% on Stanford-Dogs, and 99.7% on Oxford Flowers-101—are above prior published results, and the learned hyperedges visually align with object parts without part annotations. If the claim holds, it shows that high-order token relations and hierarchy-aware embeddings are useful for separating visually similar categories. The main caveat, visible in the experimental setup, is that comparisons use published numbers from models trained under different backbones, resolutions, and recipes.

Core claim

The paper's central claim is that fine-grained visual categories can be separated more cleanly when a model first groups image tokens into semantic regions and then organizes those regions into a hierarchy. The Semantic-Aware Aggregation Module (SAAM) does the grouping: multi-scale context is pooled into M learned prototype vectors, each prototype anchors one hyperedge of a weighted hypergraph, and every token joins every hyperedge with a soft participation weight computed by scaled dot-product similarity. Two rounds of hypergraph message passing—tokens to hyperedges, then hyperedges back to tokens—refine token features so that each token carries region-level context. The Hyperbolic Hierarch

What carries the argument

The load-bearing object is the dynamically constructed weighted hypergraph. Unlike a graph edge, which connects exactly two tokens, a hyperedge connects many tokens at once through a learned participation matrix, so one hyperedge can capture a whole semantic region such as a beak or wing. The hyperedge centroids are learnable prototypes seeded by multi-scale context vectors from the Context Generation Module, and the incidence weights are soft token-to-prototype similarities. Two hypergraph convolutions—node-to-hyperedge, then hyperedge-to-node—propagate region-level semantics back into token features. The second mechanism is the hierarchical contrastive loss: region features are merged by s

Load-bearing premise

The reported gains assume that published numbers from other models are directly comparable despite different backbones, input sizes, and training recipes, and that hyperparameters M=16 and K=0.1, chosen from accuracy curves on the same test sets used for the final tables, do not inflate the results.

What would settle it

Re-run H3Former and the closest baselines under one shared protocol—same Swin-B backbone, 448x448 input, same pretraining and training schedule—and report the gap; or hold out a validation split for choosing M and K before touching the test sets. If the margin over matched baselines falls below a standard deviation of the runs, or if test-set-tuned hyperparameters do not reproduce on held-out data, the central state-of-the-art claim would not survive.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If H3Former's results hold, fine-grained classifiers can move from selecting individual informative tokens to aggregating them into coherent semantic regions, which should improve tolerance to pose variation and background clutter.
  • The learned hyperedges act as soft part detectors: without part-level labels, each hyperedge aligns with a distinct object part, giving a path to interpretable region localization as a by-product of classification.
  • The hierarchical contrastive loss implies that imposing tree-like semantic structure in a non-Euclidean embedding can tighten intra-class clusters and widen inter-class separation beyond what Euclidean contrastive learning achieves.
  • Ablations showing that SAAM and HHCL each help, and that replacing them with pairwise GNNs or simpler hyperbolic losses hurts, support the claim that both high-order aggregation and hierarchy-aware supervision contribute to the reported accuracy.
  • The framework is backbone-agnostic in principle, since SAAM operates on multi-stage transformer features; the same token-to-region aggregation could transfer to other fine-grained recognition settings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The +4.7% gap on Stanford-Dogs is against ViT-B models, while H3Former uses Swin-B at 448x448; a matched-protocol rerun with identical pretraining, resolution, and training schedule could shrink the reported margins, so the architectural contribution may be smaller than the tables suggest.
  • Because the paper selects M=16 and K=0.1 from accuracy curves on the same test sets used for the final numbers, a validation-based tuning study would separate the method's contribution from the tuning procedure.
  • The dynamic hypergraph idea is not tied to classification: the same token-to-region aggregation with learned hyperedge prototypes could be applied to detection, segmentation, or retrieval, where high-order part relations are equally relevant. The paper does not test this.
  • The visualizations imply a specific, testable property: if hyperedges encode class-discriminative parts, then ablating or occluding the tokens of one hyperedge should produce a targeted drop in classification confidence. The paper does not perform that lesion, but it could be done with the trained model.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. H3Former is a fine-grained visual classification (FGVC) framework built on a Swin-B transformer. It proposes two main components: (i) the Semantic-Aware Aggregation Module (SAAM), which uses multi-scale context to learn hyperedge prototypes and constructs a soft hypergraph over image tokens, then performs token-to-region message passing to form region-level representations; and (ii) the Hyperbolic Hierarchical Contrastive Loss (HHCL), which applies contrastive and hierarchy-preserving constraints in Euclidean and Lorentz hyperbolic spaces. The paper reports state-of-the-art results on CUB-200-2011 (92.7%), NA-Birds (91.6%), Stanford-Dogs (95.8%), and Oxford Flowers-101 (99.7%), with ablations, visualizations, and a computational cost table.

Significance. The conceptual contribution is appealing: hypergraph-based token aggregation is a natural way to obtain semantic regions without part annotations, and combining it with a hierarchical hyperbolic loss is a plausible and timely idea. If the empirical claims are correct, H3Former would be a strong new entry in FGVC. However, the current evidence is not yet convincing: the SOTA comparisons are mostly against models with different backbones, input resolutions, and training recipes, and the HHCL ablation is inconsistent with the method equations. The paper is significant as a proposal, but its central empirical claim needs to be either substantiated with controlled experiments or appropriately qualified.

major comments (4)
  1. [Sec. IV-C, Tables I-IV] The claim that H3Former 'consistently achieves the highest accuracy across all datasets' is not supported by controlled comparisons. The proposed method uses Swin-B at 448x448, pretrained on ImageNet-22K (or ImageNet-1K for Stanford-Dogs), while most baselines use ViT-B, Xception, DenseNet, or ResNet at lower resolutions with different recipes. The only same-backbone comparison is TransIFC+ on NA-Birds (+0.7%). The margins of +1.0% over IELT on CUB, +4.7% over FAL-ViT on Stanford-Dogs, and +0.7% over I2-HOFI on Flowers-101 could plausibly reflect the backbone/setup rather than the proposed modules. Please report re-runs of the strongest baselines under the same protocol, or at least compare against a Swin-B/448 trained with the same augmentations and pretraining, and temper the SOTA claim accordingly.
  2. [Sec. IV-E, Fig. 8] The hyperparameters M=16 and K=0.1 are selected from accuracy curves shown in Fig. 8. The text does not state whether these curves are on a held-out validation split or on the same test sets used for the final reported numbers. Since the main tables use these selected values, the headline results may be optimistic selections rather than independent evaluations. Please clarify the split used for tuning, and report means and standard deviations over multiple runs (or state the single-run variance) to support the reported margins.
  3. [Sec. III-C, Eq. (13)-(16) vs Sec. IV-E, Table VI] The loss definition is inconsistent. The method section defines a single supervised contrastive loss L_con (Eq. 14) that combines Euclidean and hyperbolic distances, plus L_hpop (Eq. 15), giving L_HHCL = L_con + β L_hpop (Eq. 16). Table VI, however, ablates three separate components: L_hcon, L_econ, and L_hpop. The losses L_hcon and L_econ are never defined in Sec. III-C, and the text in Sec. IV-E describes L_econ as an 'entropy minimization' regularizer in hyperbolic space, which is not present in the equations. This means the ablation that attributes the final gain to HHCL cannot be checked. Please either define L_hcon and L_econ explicitly in the method, or align Table VI with Eq. (16) by ablating L_con and L_hpop.
  4. [Sec. III-C, Eq. (15)] The Hypergraph Partial Order Preservation Loss is stated to enforce that higher-level features lie closer to their children, but Eq. (15) simply penalizes the hyperbolic distance d_L(H^{ℓ+1}, H^ℓ) with a ReLU, which is redundant because a distance is already non-negative. There is no margin or asymmetric constraint, and no explicit notion of an order such as an increasing norm or time coordinate. As written, this term could be trivially minimized by collapsing all levels to a single point, which would not preserve a hierarchy. A margin-based or norm-based ordering constraint would be needed to justify the claim that Eq. (15) preserves a partial order.
minor comments (6)
  1. [Sec. III-B] Typo: 'interpolateed' should be 'interpolated'.
  2. [Sec. IV-E, Table VI caption] The caption reads 'DIFFERENT RADIOS OF LOSSES'; 'radios' should be 'ratios'.
  3. [Fig. 8 caption] The caption reads 'he curvature K'; should be 'the curvature K'.
  4. [Table VIII] The caption states that 'All methods are evaluated under the same input resolution to ensure a fair comparison,' but the table lists input sizes 448 for ViT-based models, 384 for Swin-Base, and both 384 and 448 for Ours. Please clarify what 'same input resolution' means here.
  5. [Eq. (15)] The ReLU around a non-negative distance is unnecessary; if a margin is intended, it should be written explicitly.
  6. [Eq. (5)] The affinity computation uses a softmax without a temperature scale. Given the later choice of temperature τ=0.1 for the contrastive loss, the absence of a temperature in Eq. (5) may produce very peaked or flat assignments; please comment or add a temperature if needed.

Circularity Check

0 steps flagged

No circular derivation; score reflects a minor, non-load-bearing self-citation and protocol concerns, not a reduction of the central claim to its inputs.

full rationale

The paper's derivation chain is not circular. SAAM constructs a dynamic weighted hypergraph from token features and learnable prototypes (Eqs. 3-8) and HHCL adds a supervised contrastive regularizer with a parent-child consistency term (Eqs. 13-16); both are trained end-to-end with cross-entropy, which is standard supervised learning rather than a self-referential derivation. No fitted parameter is relabeled as a prediction: the reported accuracies are measured after training, not derived from the method's equations by construction. The only self-citation of note is SoftHGNN [63], used as an ablation baseline in Table VII; it shares co-authors with the present paper but is not load-bearing for the central SOTA claim, so it warrants only the rubric's minor-self-citation score. Two non-circular weaknesses should be flagged: (1) Fig. 8 selects hyperparameters M and K using accuracy on the same benchmark test sets later reported as final results, with no validation split described, so the reported margins may be optimistically biased; (2) Sec. IV-E and Table VI refer to separate losses L_hcon and L_econ, but Eqs. 13-16 define only a hybrid L_con and L_hpop, making the ablation attribution to HHCL components unverifiable. These are reproducibility and evaluation-protocol concerns, not circularity: the central architectural claim does not reduce to any equation's own input.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 1 invented entities

The central claim rests on the correctness of the hypergraph assignment (trained end-to-end), the chosen hyperparameters, and the comparability of benchmark numbers. The method introduces one latent construct (learnable semantic prototypes) and several hand-tuned constants; no invented physical entities.

free parameters (7)
  • M (number of hyperedges) = 16
    Number of semantic region clusters; chosen by ablation over {2,4,8,16,32} in Fig. 8; affects all downstream region features and accuracy.
  • K (Lorentz curvature) = 0.1
    Curvature of hyperbolic embedding; selected from {0.05,0.1,0.2,0.5} in Fig. 8; central to HHCL distance.
  • τ (temperature) = 0.1
    Temperature in supervised contrastive loss Eq. (14); set by hand without ablation.
  • λ (Euclidean/hyperbolic balance) = 1.0
    Weight between Euclidean and hyperbolic distance in Eq. (13); set by hand.
  • β (hierarchy preservation weight) = 0.1
    Weight for L_hpop in Eq. (16); tuned; at 0.5 or 0.2 performance drops per Tab. VI.
  • α (overall HHCL weight) = not reported
    Appears in Eq. (1) but no value is explicitly reported; ablation Table VI implies α=1.0 with per-component weights.
  • Hierarchy fusion ratios = {16,8,4,1}
    Region fusion ratios across four hierarchy levels; chosen by hand; determine the tree structure used by HHCL.
axioms (5)
  • domain assumption Pretrained Swin Transformer features (ImageNet-22K/1K) transfer to fine-grained domains.
    Used as backbone in Sec III-A and IV-B; no experiments without pretraining are reported.
  • domain assumption Published baseline accuracies are directly comparable despite different backbones, input sizes, and training recipes.
    Comparison tables in Sec IV-C mix ViT-B, Swin-B, DenseNet, Xception baselines; the paper assumes protocol differences do not dominate architectural gains.
  • ad hoc to paper Learnable prototype plus multi-scale context yields semantically coherent hyperedges that are discriminative for FGVC.
    Central design hypothesis of SAAM (Eq. 3-5); the paper provides visual evidence but no formal guarantee.
  • ad hoc to paper Merging region features by similarity produces a valid category hierarchy for the partial-order loss.
    Eq. (12) assumes similarity-based aggregation operator A(·) yields meaningful parent-child structure; never validated against human-defined hierarchies.
  • domain assumption The Lorentz model with fixed K=0.1 expresses fine-grained category hierarchy better than Euclidean geometry.
    Section III-C motivation; no comparison with other K or other non-Euclidean models except a small ablation.
invented entities (1)
  • Semantic hyperedge prototypes P_m no independent evidence
    purpose: Learnable centroids that define hyperedges and aggregate tokens into semantic regions (Eq. 3).
    Internal parameters; the only evidence is in-paper visualization of hyperedge activations; no external prediction or falsifiable handle.

pith-pipeline@v1.3.0-alltime-deepseek · 17788 in / 12444 out tokens · 111403 ms · 2026-08-03T22:26:19.879787+00:00 · methodology

0 comments
read the original abstract

Fine-Grained Visual Classification (FGVC) remains a challenging task due to subtle inter-class differences and large intra-class variations. Existing approaches typically rely on feature-selection mechanisms or region-proposal strategies to localize discriminative regions for semantic analysis. However, these methods often fail to capture discriminative cues comprehensively while introducing substantial category-agnostic redundancy. To address these limitations, we propose H3Former, a novel token-to-region framework that leverages high-order semantic relations to aggregate local fine-grained representations with structured region-level modeling. Specifically, we propose the Semantic-Aware Aggregation Module (SAAM), which exploits multi-scale contextual cues to dynamically construct a weighted hypergraph among tokens. By applying hypergraph convolution, SAAM captures high-order semantic dependencies and progressively aggregates token features into compact region-level representations. Furthermore, we introduce the Hyperbolic Hierarchical Contrastive Loss (HHCL), which enforces hierarchical semantic constraints in a non-Euclidean embedding space. The HHCL enhances inter-class separability and intra-class consistency while preserving the intrinsic hierarchical relationships among fine-grained categories. Comprehensive experiments conducted on four standard FGVC benchmarks validate the superiority of our H3Former framework.

Figures

Figures reproduced from arXiv: 2511.10260 by Kuiyang Huang, Siqi Li, Yongji Zhang, Yue Gao, Yu Jiang.

Figure 1
Figure 1. Figure 1: Hyperedges (E 1–E 4 ) of hypergraph H = (V, E) generated by our H3Former. Distinct hyperedges correspond to meaningful semantic regions, e.g., tail feathers, wing, beak, and eye. The learned hypergraphs automatically highlight key discriminative parts without any part-level su￾pervision. H3Former adaptively constructs coherent semantic regions through its hypergraph construction mechanism, bridging local t… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of different FGVC paradigms. (a) Feature-selection based methods perform token filtering in the feature space to retain features most relevant to fine-grained recognition, but overlook coherent semantic structure. (b) Region-relation based methods learn pairwise dependencies among predefined regions, typically obtained from RPNs, which may introduce redundant and category-agnostic information.… view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the proposed H3Former framework. The Semantic-Aware Aggregation Module (SAAM) constructs a weighted hypergraph to capture high-order semantic relations and progressively aggregates tokens into semantically coherent regions. Meanwhile, the Hyperbolic Hierarchical Contrastive Loss (HHCL) operates on the resulting hierarchical region representations to enforce fine-grained category separation and … view at source ↗
Figure 4
Figure 4. Figure 4: The architecture of the Context Generation Module (CGM). The CGM utilizes the token features and attention maps from each stage to generate corresponding context vectors that encode multi-scale contextual information. When window-based attention is used, the attention maps are processed along the dashed path to produce the importance vector, which reflects the relative significance of tokens within each wi… view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of hierarchical hypergraph modeling and HHCL loss. (a) HHCL consists of Lhpop for hierarchical consistency, Lhcon for hyperbolic contrastive learning, and Lecon for euclidean discrimination. (b) Region-level features are hierarchically merged based on semantic similarity. (c) SAAM performs soft hypergraph message passing from tokens to regions and back. a unified representation that connects t… view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of hyperedges in H3Former. Each row shows two input images from same dataset and the activation maps of four hyperedges. Each hyperedge captures a distinct semantic region, e.g., the beak, wings, or feet of the bird. This illustrates the semantic-aware and complementary nature of our H3Former. structuring capability of hypergraphs. More importantly, our model outperforms recent strong fine-gr… view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of hyperedges learned with different loss designs. Each column corresponds to one hyperedge, and each row shows the token￾level activation maps of the models from Tab. VII (rows 3 and 4) and our proposed method. Replacing our HHCL with alternative hyperbolic loss functions leads to less compact or inconsistent semantic grouping, while our HHCL tightly aligns with the hypergraph structure, pro… view at source ↗
Figure 8
Figure 8. Figure 8: Influence of hyperparameters on classification accuracy on the CUB-200-2011 and Stanford-Dogs datasets. (a) Accuracy curves with the numbers of hyperedges M. (b) Accuracy curves with he curvature K in the Lorentzian embedding. (a) Backbone with HHCL (b) Backbone with SAAM (c) Ours [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: t-SNE visualizations on the Stanford-Dogs dataset. (a) Features with HHCL only. (b) Features with SAAM only. (c) Features incorporating both SAAM and HHCL, demonstrates clearer clustering and enhanced inter￾class separability. a single geometric constraint. This suggests that while each component is beneficial, their contributions are most effective when balanced, reflecting their complementary roles in st… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

64 extracted references · 6 linked inside Pith

  1. [1]

    A Survey on Vision Transformer,

    K. Han, Y . Wang, H. Chen, X. Chen, J. Guo, Z. Liu, Y . Tang, A. Xiao, C. Xu, Y . Xuet al., “A Survey on Vision Transformer,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 1, pp. 87–110, 2022

    2022

  2. [2]

    Fine-Grained Image Analysis with Deep Learning: A Survey,

    X.-S. Wei, Y .-Z. Song, O. Mac Aodha, J. Wu, Y . Peng, J. Tang, J. Yang, and S. Belongie, “Fine-Grained Image Analysis with Deep Learning: A Survey,”IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 44, no. 12, pp. 8927–8948, 2021

    2021

  3. [3]

    A Survey on Deep Learning- based Fine-Grained Object Classification and Semantic Segmentation,

    B. Zhao, J. Feng, X. Wu, and S. Yan, “A Survey on Deep Learning- based Fine-Grained Object Classification and Semantic Segmentation,” International Journal of Automation and Computing, vol. 14, no. 2, pp. 119–135, 2017

    2017

  4. [4]

    Transformers in Vision: A Survey,

    S. Khan, M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah, “Transformers in Vision: A Survey,”ACM computing surveys (CSUR), vol. 54, no. 10s, pp. 1–41, 2022

    2022

  5. [5]

    A Vision Transformer for Fine-Grained Classification by Reducing Noise and Enhancing Discriminative Information,

    Z.-C. Zhang, Z.-D. Chen, Y . Wang, X. Luo, and X.-S. Xu, “A Vision Transformer for Fine-Grained Classification by Reducing Noise and Enhancing Discriminative Information,”Pattern Recognition, vol. 145, p. 109979, 2024

    2024

  6. [6]

    Multi-Part Token Transformer with Dual Contrastive Learning for Fine-Grained Image Classification,

    C. Wang, H. Fu, and H. Ma, “Multi-Part Token Transformer with Dual Contrastive Learning for Fine-Grained Image Classification,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 7648–7656

    2023

  7. [7]

    Feature Fusion Vision Transformer for Fine-Grained Visual Categorization,

    J. Wang, X. Yu, and Y . Gao, “Feature Fusion Vision Transformer for Fine-Grained Visual Categorization,”arXiv preprint arXiv:2107.02341, 2021

    Pith/arXiv arXiv 2021

  8. [8]

    TransFG: A Transformer Architecture for Fine-Grained Recognition,

    J. He, J.-N. Chen, S. Liu, A. Kortylewski, C. Yang, Y . Bai, and C. Wang, “TransFG: A Transformer Architecture for Fine-Grained Recognition,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 1, 2022, pp. 852–860

    2022

  9. [9]

    Fine-Grained Visual Classifica- tion via Internal Ensemble Learning Transformer,

    Q. Xu, J. Wang, B. Jiang, and B. Luo, “Fine-Grained Visual Classifica- tion via Internal Ensemble Learning Transformer,”IEEE Transactions on Multimedia, 2023

    2023

  10. [10]

    Local and Global Logit Adjustments for Long-Tailed Learning,

    Y . Tao, J. Sun, H. Yang, L. Chen, X. Wang, W. Yang, D. Du, and M. Zheng, “Local and Global Logit Adjustments for Long-Tailed Learning,” inProceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 11 783–11 792

    2023

  11. [11]

    SR-GNN: Spatial Relation-Aware Graph Neural Network for Fine-Grained Image Categorization,

    A. Bera, Z. Wharton, Y . Liu, N. Bessis, and A. Behera, “SR-GNN: Spatial Relation-Aware Graph Neural Network for Fine-Grained Image Categorization,”IEEE Transactions on Image Processing, vol. 31, pp. 6017–6031, 2022

    2022

  12. [12]

    Interweaving Insights: High-Order Feature Interaction for Fine-Grained Visual Recognition,

    A. Sikdar, Y . Liu, S. Kedarisetty, Y . Zhao, A. Ahmed, and A. Behera, “Interweaving Insights: High-Order Feature Interaction for Fine-Grained Visual Recognition,”International Journal of Computer Vision, vol. 133, no. 4, pp. 1755–1779, 2025

    2025

  13. [13]

    Hypergraph-Induced Convolutional Networks for Visual Classifica- tion,

    H. Shi, Y . Zhang, Z. Zhang, N. Ma, X. Zhao, Y . Gao, and J. Sun, “Hypergraph-Induced Convolutional Networks for Visual Classifica- tion,”IEEE transactions on neural networks and learning systems, vol. 30, no. 10, pp. 2963–2972, 2018

    2018

  14. [14]

    Adaptive Neural Message Passing for Inductive Learning on Hypergraphs,

    D. Arya, D. K. Gupta, S. Rudinac, and M. Worring, “Adaptive Neural Message Passing for Inductive Learning on Hypergraphs,”IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024

    2024

  15. [15]

    HGFormer: Topology-Aware Vision Transformer with Hypergraph Learning,

    H. Wang, S. Zhang, and B. Leng, “HGFormer: Topology-Aware Vision Transformer with Hypergraph Learning,”IEEE Transactions on Multi- media, 2025

    2025

  16. [16]

    The Caltech-Ucsd Birds-200-2011 Dataset,

    C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie, “The Caltech-Ucsd Birds-200-2011 Dataset,” 2011

    2011

  17. [17]

    Novel Dataset for Fine-Grained Image Categorization: Stanford Dogs,

    A. Khosla, N. Jayadevaprakash, B. Yao, and L. Fei-Fei, “Novel Dataset for Fine-Grained Image Categorization: Stanford Dogs,” inProceedings of the CVPR Workshop on Fine-Grained Visual Categorization (FGVC), vol. 2. Citeseer, 2011, pp. 1–2

    2011

  18. [18]

    Attention is All You Need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is All You Need,”Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  19. [19]

    Rams- trans: Recurrent attention multi-scale transformer for fine-grained image recognition,

    Y . Hu, X. Jin, Y . Zhang, H. Hong, J. Zhang, Y . He, and H. Xue, “Rams- trans: Recurrent attention multi-scale transformer for fine-grained image recognition,” inProceedings of the 29th ACM International Conference on Multimedia, 2021, pp. 4239–4248

    2021

  20. [20]

    Delving into Multimodal Prompting for Fine-Grained Visual Classification,

    X. Jiang, H. Tang, J. Gao, X. Du, S. He, and Z. Li, “Delving into Multimodal Prompting for Fine-Grained Visual Classification,” inPro- ceedings of the AAAI conference on artificial intelligence, vol. 38, no. 3, 2024, pp. 2570–2578

    2024

  21. [21]

    Fine-Grained Recognition without Part Annotations,

    J. Krause, H. Jin, J. Yang, and L. Fei-Fei, “Fine-Grained Recognition without Part Annotations,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 5546–5555

    2015

  22. [22]

    Bird Species Categorization using Pose Normalized Deep Convolutional Nets,

    S. Branson, G. Van Horn, S. Belongie, and P. Perona, “Bird Species Categorization using Pose Normalized Deep Convolutional Nets,”arXiv preprint arXiv:1406.2952, 2014

    Pith/arXiv arXiv 2014

  23. [23]

    Fully Convolutional Attention Networks for Fine-Grained Recognition,

    X. Liu, T. Xia, J. Wang, Y . Yang, F. Zhou, and Y . Lin, “Fully Convolutional Attention Networks for Fine-Grained Recognition,”arXiv preprint arXiv:1603.06765, 2016

    Pith/arXiv arXiv 2016

  24. [24]

    Bilinear CNN Models for Fine-Grained Visual Recognition,

    T.-Y . Lin, A. RoyChowdhury, and S. Maji, “Bilinear CNN Models for Fine-Grained Visual Recognition,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 1449–1457

    2015

  25. [25]

    The Devil is in the Channels: Mutual-Channel Loss for Fine-Grained Image Classification,

    D. Chang, Y . Ding, J. Xie, A. K. Bhunia, X. Li, Z. Ma, M. Wu, J. Guo, and Y .-Z. Song, “The Devil is in the Channels: Mutual-Channel Loss for Fine-Grained Image Classification,”IEEE Transactions on Image Processing, vol. 29, pp. 4683–4695, 2020

    2020

  26. [26]

    Cross-Part Learning for Fine-Grained Image Classification,

    M. Liu, C. Zhang, H. Bai, R. Zhang, and Y . Zhao, “Cross-Part Learning for Fine-Grained Image Classification,”IEEE Transactions on Image Processing, vol. 31, pp. 748–758, 2021

    2021

  27. [27]

    Mask-CNN: Localizing Parts and Selecting Descriptors for Fine-Grained Bird Species Catego- rization,

    X.-S. Wei, C.-W. Xie, J. Wu, and C. Shen, “Mask-CNN: Localizing Parts and Selecting Descriptors for Fine-Grained Bird Species Catego- rization,”Pattern Recognition, vol. 76, pp. 704–714, 2018

    2018

  28. [28]

    Part-Guided Relational Trans- formers for Fine-Grained Visual Recognition,

    Y . Zhao, J. Li, X. Chen, and Y . Tian, “Part-Guided Relational Trans- formers for Fine-Grained Visual Recognition,”IEEE Transactions on Image Processing, vol. 30, pp. 9470–9481, 2021

    2021

  29. [29]

    Weakly Supervised Posture Min- ing for Fine-Grained Classification,

    Z. Tang, H. Yang, and C. Y .-C. Chen, “Weakly Supervised Posture Min- ing for Fine-Grained Classification,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 23 735–23 744

    2023

  30. [30]

    Improving Visual Recognition with Hyperbolical Visual Hierarchy Mapping,

    H. Kwon, J. Jang, J. Kim, K. Kim, and K. Sohn, “Improving Visual Recognition with Hyperbolical Visual Hierarchy Mapping,” inProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 17 364–17 374

    2024

  31. [31]

    Mode Hypergraph Neural Network,

    S. Ji, Y . Feng, D. Di, S. Ying, and Y . Gao, “Mode Hypergraph Neural Network,”IEEE Transactions on Neural Networks and Learning Systems, 2025

    2025

  32. [32]

    LightHGNN: Distilling Hyper- graph Neural Networks into MLPs for 100x Faster Inference,

    Y . Feng, Y . Luo, S. Ying, and Y . Gao, “LightHGNN: Distilling Hyper- graph Neural Networks into MLPs for 100x Faster Inference,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  33. [33]

    Hgnn+: General Hypergraph Neural Networks,

    Y . Gao, Y . Feng, S. Ji, and R. Ji, “Hgnn+: General Hypergraph Neural Networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3181–3199, 2022

    2022

  34. [34]

    Hyper-Yolo: When Visual Object Detection Meets Hypergraph Computation,

    Y . Feng, J. Huang, S. Du, S. Ying, J.-H. Yong, Y . Li, G. Ding, R. Ji, and Y . Gao, “Hyper-Yolo: When Visual Object Detection Meets Hypergraph Computation,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  35. [35]

    Vision Hgnn: An Image Is More Than A Graph of Nodes,

    Y . Han, P. Wang, S. Kundu, Y . Ding, and Z. Wang, “Vision Hgnn: An Image Is More Than A Graph of Nodes,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19 878–19 888

    2023

  36. [36]

    Hyperbolic Category Discovery,

    Y . Liu, Z. He, and K. Han, “Hyperbolic Category Discovery,” inPro- ceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 9891–9900

    2025

  37. [37]

    HLFormer: Enhancing Partially Rel- evant Video Retrieval with Hyperbolic Learning,

    L. Jun, W. Jinpeng, T. Chaolei, L. Niu, C. Long, Z. Min, W. Yaowei, X. Shu-Tao, and C. Bin, “HLFormer: Enhancing Partially Rel- evant Video Retrieval with Hyperbolic Learning,”arXiv preprint arXiv:2507.17402, 2025

    Pith/arXiv arXiv 2025

  38. [38]

    Building A Bird Recognition APP and Large Scale Dataset with Citizen Scientists: The Fine Print in Fine- Grained Dataset Collection,

    G. Van Horn, S. Branson, R. Farrell, S. Haber, J. Barry, P. Ipeirotis, P. Perona, and S. Belongie, “Building A Bird Recognition APP and Large Scale Dataset with Citizen Scientists: The Fine Print in Fine- Grained Dataset Collection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2015, pp. 595–604

    2015

  39. [39]

    Automated Flower Classification Over A Large Number of Classes,

    M.-E. Nilsback and A. Zisserman, “Automated Flower Classification Over A Large Number of Classes,” in2008 Sixth Indian conference on computer vision, graphics & image processing. IEEE, 2008, pp. 722–729

    2008

  40. [40]

    Swin Transformer: Hierarchical Vision Transformer using Shifted Win- dows,

    Z. Liu, Y . Lin, Y . Cao, H. Hu, Y . Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical Vision Transformer using Shifted Win- dows,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 10 012–10 022

    2021

  41. [41]

    ImageNet-21k Pretraining for The Masses,

    T. Ridnik, E. Ben-Baruch, A. Noy, and L. Zelnik-Manor, “ImageNet-21k Pretraining for The Masses,”arXiv preprint arXiv:2104.10972, 2021

    Pith/arXiv arXiv 2021

  42. [42]

    ImageNet: A Large-Scale Hierarchical Image Database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A Large-Scale Hierarchical Image Database,” in2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255

    2009

  43. [43]

    An Attention- Locating Algorithm for Eliminating Background Effects in Fine-Grained Visual Classification,

    Y . Huang, Z. Hechen, M. Zhou, Z. Li, and S. Kwong, “An Attention- Locating Algorithm for Eliminating Background Effects in Fine-Grained Visual Classification,”IEEE Transactions on Circuits and Systems for Video Technology, 2025. 12

    2025

  44. [44]

    Filtration and Distillation: Enhancing Region Attention for Fine-Grained Visual Categorization,

    C. Liu, H. Xie, Z.-J. Zha, L. Ma, L. Yu, and Y . Zhang, “Filtration and Distillation: Enhancing Region Attention for Fine-Grained Visual Categorization,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 34, no. 07, 2020, pp. 11 555–11 562

    2020

  45. [45]

    Fine-Grained Visual Categorization by Localizing Object Parts with Single Image,

    X. Zheng, L. Qi, Y . Ren, and X. Lu, “Fine-Grained Visual Categorization by Localizing Object Parts with Single Image,”IEEE Transactions on Multimedia, vol. 23, pp. 1187–1199, 2020

    2020

  46. [46]

    AP-CNN: Weakly Supervised Attention Pyramid Convolutional Neural Network for Fine-Grained Visual Classification,

    Y . Ding, Z. Ma, S. Wen, J. Xie, D. Chang, Z. Si, M. Wu, and H. Ling, “AP-CNN: Weakly Supervised Attention Pyramid Convolutional Neural Network for Fine-Grained Visual Classification,”IEEE Transactions on Image Processing, vol. 30, pp. 2826–2836, 2021

    2021

  47. [47]

    Fine-Grained Object Classification via Self-Supervised Pose Alignment,

    X. Yang, Y . Wang, K. Chen, Y . Xu, and Y . Tian, “Fine-Grained Object Classification via Self-Supervised Pose Alignment,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 7399–7408

    2022

  48. [48]

    Granularity-Aware Distillation and Structure Modeling Region Proposal Network for Fine- Grained Image classification,

    X. Ke, Y . Cai, B. Chen, H. Liu, and W. Guo, “Granularity-Aware Distillation and Structure Modeling Region Proposal Network for Fine- Grained Image classification,”Pattern Recognition, vol. 137, p. 109305, 2023

    2023

  49. [49]

    Learning Mutually Exclusive Part Rep- resentations for Fine-Grained Image Classification,

    C. Wang, H. Fu, and H. Ma, “Learning Mutually Exclusive Part Rep- resentations for Fine-Grained Image Classification,”IEEE Transactions on Multimedia, vol. 26, pp. 3113–3124, 2023

    2023

  50. [50]

    Vit-Net: Interpretable Vision Transform- ers with Neural Tree Decoder,

    S. Kim, J. Nam, and B. C. Ko, “Vit-Net: Interpretable Vision Transform- ers with Neural Tree Decoder,” inInternational conference on machine learning. PMLR, 2022, pp. 11 162–11 172

    2022

  51. [51]

    TransIFC: Invariant Cues-Aware Feature Concentration Learning for Efficient Fine- Grained Bird Image Classification,

    H. Liu, C. Zhang, Y . Deng, B. Xie, T. Liu, and Y .-F. Li, “TransIFC: Invariant Cues-Aware Feature Concentration Learning for Efficient Fine- Grained Bird Image Classification,”IEEE Transactions on Multimedia, vol. 27, pp. 1677–1690, 2023

    2023

  52. [52]

    Learning Attentive Pairwise In- teraction for Fine-Grained Classification,

    P. Zhuang, Y . Wang, and Y . Qiao, “Learning Attentive Pairwise In- teraction for Fine-Grained Classification,” inProceedings of the AAAI conference on artificial intelligence, vol. 34, no. 07, 2020, pp. 13 130– 13 137

    2020

  53. [53]

    Context-Aware Attentional Pooling for Fine-Grained Visual Classification,

    A. Behera, Z. Wharton, P. R. Hewage, and A. Bera, “Context-Aware Attentional Pooling for Fine-Grained Visual Classification,” inProceed- ings of the AAAI conference on artificial intelligence, vol. 35, no. 2, 2021, pp. 929–937

    2021

  54. [54]

    Counterfactual Attention Learning for Fine-Grained Visual Categorization and Re-Identification,

    Y . Rao, G. Chen, J. Lu, and J. Zhou, “Counterfactual Attention Learning for Fine-Grained Visual Categorization and Re-Identification,” inPro- ceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 1025–1034

    2021

  55. [55]

    PBC: Polygon- based Classifier for Fine-Grained Categorization,

    C. Huang, H. Li, Y . Xie, Q. Wu, and B. Luo, “PBC: Polygon- based Classifier for Fine-Grained Categorization,”IEEE Transactions on Multimedia, vol. 19, no. 4, pp. 673–684, 2016

    2016

  56. [56]

    Interactive: Inter- Layer Activeness Propagation,

    L. Xie, L. Zheng, J. Wang, A. L. Yuille, and Q. Tian, “Interactive: Inter- Layer Activeness Propagation,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 270–279

    2016

  57. [57]

    Borrowing Treasures from The Wealthy: Deep Transfer Learning Through Selective Joint Fine-Tuning,

    W. Ge and Y . Yu, “Borrowing Treasures from The Wealthy: Deep Transfer Learning Through Selective Joint Fine-Tuning,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 1086–1095

    2017

  58. [58]

    Object-Part Attention Model for Fine- Grained Image Classification,

    Y . Peng, X. He, and J. Zhao, “Object-Part Attention Model for Fine- Grained Image Classification,”IEEE Transactions on Image Processing, vol. 27, no. 3, pp. 1487–1500, 2017

    2017

  59. [59]

    Large scale Fine-Grained Categorization and Domain-Specific Transfer Learning,

    Y . Cui, Y . Song, C. Sun, A. Howard, and S. Belongie, “Large scale Fine-Grained Categorization and Domain-Specific Transfer Learning,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 4109–4118

    2018

  60. [60]

    Learning A Mixture of Granularity-Specific Experts for Fine-Grained Categorization,

    L. Zhang, S. Huang, W. Liu, and D. Tao, “Learning A Mixture of Granularity-Specific Experts for Fine-Grained Categorization,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 8331–8340

    2019

  61. [61]

    Deep Learning on Small Datasets without Pre- training Using Cosine Loss,

    B. Barz and J. Denzler, “Deep Learning on Small Datasets without Pre- training Using Cosine Loss,” inProceedings of the IEEE/CVF winter conference on applications of computer vision, 2020, pp. 1371–1380

    2020

  62. [62]

    Bi-Modal Progres- sive Mask Attention for Fine-Grained Recognition,

    K. Song, X.-S. Wei, X. Shu, R.-J. Song, and J. Lu, “Bi-Modal Progres- sive Mask Attention for Fine-Grained Recognition,”IEEE Transactions on Image Processing, vol. 29, pp. 7006–7018, 2020

    2020

  63. [63]

    SoftHGNN: Soft Hypergraph Neural Networks for General Visual Recognition,

    M. Lei, Y . Wu, S. Li, X. Zheng, J. Wang, Y . Gao, and S. Du, “SoftHGNN: Soft Hypergraph Neural Networks for General Visual Recognition,”arXiv preprint arXiv:2505.15325, 2025

    Pith/arXiv arXiv 2025

  64. [64]

    Understanding Hyperbolic Metric Learning through Hard Negative Sampling,

    Y . Yue, F. Lin, G. Mou, and Z. Zhang, “Understanding Hyperbolic Metric Learning through Hard Negative Sampling,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 1891–1903. Yongji Zhangis currently working toward the Ph.D. degree in the College of Computer Science and Technology, Jilin University, China. He re...

    2024