Pith. sign in

REVIEW 2 major objections 5 minor 32 references

FlexiGrad: Adaptive Gradient Modulation for Hierarchical Fine-Grained Classification

T0 review · 2 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read FlexiGrad, a parameter-free gradient-modulation rule, removes only the conflicting coarse–fine component and reinforces partial agreement, lifting multi-granularity accuracy on three fine-grained benchmarks.

desk verdict PCGrad plus a cosine-weighted cooperation branch; consistent but modest gains, yet the update rule is ambiguous and the theory is a non sequitur. read the letter →

arxiv 2607.17563 v2 pith:C6K6H5XX submitted 2026-07-20 cs.CV

Zilu Zhou , Dongliang Chang , Junhan Chen , Zhanyu Ma This is my paper
classification cs.CV
keywords multi-granularityclassificationfine-grainedvisualhierarchicallabelsgradientconflictsurgeryparameter-freecoarse-to-finelearningmulti-task
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 argues that the instability of joint coarse-and-fine training comes from gradient conflicts on the shared backbone, not from the labels themselves. It introduces FlexiGrad, a parameter-free backpropagation-time rule: when gradients from two granularity levels point in opposite directions, it subtracts the projection of one onto the other; when they partially agree, it blends them with a linear weight (1+cos θ)/2. The resulting update is a surrogate descent direction for the total objective, and experiments on CUB-200-2011, FGVC-Aircraft and Stanford Cars report consistent accuracy gains at every granularity, with the largest gains on the finest level and on hard taxonomic subtrees. The practical payoff is a drop-in optimisation fix that works across CNN and transformer backbones without modifying the architecture.

What carries the argument

The central mechanism is the pairwise cosine similarity cosθ_ij between the gradients of two granularity tasks. When the cosine is negative, FlexiGrad subtracts the projection of one gradient onto the other, removing only the conflicting component; when it is positive, it blends the two gradients with a smooth linear weight (1+cos θ_ij)/2, reinforcing the shared direction while preserving each task's individual component. The corrected gradients are summed to form the final parameter update, so the method intervenes purely at backpropagation time and needs no architecture changes or extra parameters.

What would settle it

On the three reported benchmarks, train with the same schedule but replace the cosine-based weights with a fixed constant w=0.5 (or with randomly shuffled pairwise cosines). If the accuracy gain over vanilla joint training matches FlexiGrad's, the alignment-sensitivity of the rule is not what drives the improvement. A second check: train with plain SGD instead of Adam; the theory only certifies descent for the aggregate gradient, so if the gain vanishes under SGD, the mechanism's causal claim would be in doubt.

Watch

Extended reading notes

Core claim

The paper's central claim is that the naive sum of coarse and fine task gradients is suboptimal because it treats all cross-level interactions as equally compatible. FlexiGrad replaces it with a hierarchy-aware update: for each pair of task gradients, on conflict (negative cosine) the projection of one onto the other is removed; on partial agreement the gradient is re-weighted as a convex combination of itself and its projection, using w_ij = (1+cos θ_ij)/2. The paper shows each corrected gradient keeps a non-negative inner product with its original, so the sum remains a descent direction for the total loss. Empirically, on three benchmarks and across four base models, this yields higher mul

Load-bearing premise

The load-bearing premise is that coarse and fine supervision are best treated as two separate tasks whose harmful interactions can be repaired by pairwise gradient projections; if the reported gains actually come from Adam-specific dynamics or from an accidental regularisation effect of the reweighting, the 'hierarchical gradient conflict' explanation would not be the cause.

Editorial extensions

If this is right

  • Joint training with FlexiGrad reports higher accuracy at every granularity than vanilla joint training, with the largest absolute gains at the finest label level (species or model).
  • FlexiGrad is a drop-in backpropagation module: no architectural change, no extra parameters, and only a small training-time overhead (about 7% in the reported CUB setting).
  • The gains persist across backbone families: the paper reports improvements for both CNN-based and transformer-based fine-grained models when FlexiGrad is plugged in.
  • The improvement is largest precisely where coarse-to-fine interference is strongest: hard families and subtrees show the biggest species-level gains, up to about 19 percentage points on CUB in the paper's analysis.
  • Because the corrected update is a surrogate descent direction for the total loss, the method can be combined with any optimiser and any hierarchical head setup without changing the objective.

Reading between the lines

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

  • The same pairwise alignment rule should extend to hierarchies deeper than two levels and to any ordered set of tasks, since the procedure is already sequential and pairwise; a testable variant would apply it to kingdom–phylum–class–order taxonomies and check whether per-level gains compound.
  • The paper's descent-direction guarantee covers only the aggregate gradient, so the method's benefit under adaptive optimisers could partly be a learning-rate interaction; running FlexiGrad with plain SGD under the same schedule would isolate the mechanism's contribution.
  • Because the weight w=(1+cos θ)/2 assumes gradients are on comparable scales, per-task loss scaling could change the outcome; a whitened-gradient variant would test whether normalisation is needed for very different loss magnitudes.
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

2 major / 5 minor

Summary. The paper proposes FlexiGrad, a parameter-free gradient-modulation method for hierarchical fine-grained classification. Given task gradients from coarse and fine classifiers, FlexiGrad measures pairwise cosine similarity, removes the conflicting component when gradients disagree, and mixes the gradient with a projected component using weight w_ij=(1+cosθ_ij)/2 when they partially agree. The corrected per-task gradients are summed and used in place of the raw multi-task gradient. Experiments on CUB-200-2011, FGVC-Aircraft, and Stanford Cars with ResNet-50, Swin-B, ViT-B/16, and several FGVC models report consistent accuracy gains over Vanilla single, PCGrad, and FGoN, with ablations and visualizations supporting the claimed behavior.

Significance. If the exact update is specified and the empirical results are reproducible, FlexiGrad would be a useful, simple, and architecture-agnostic contribution to hierarchical FGVC. It is genuinely parameter-free apart from a small epsilon, it does not alter the network structure, and the reported gains are consistent across datasets and backbones. The core idea of treating coarse-to-fine gradient interactions asymmetrically and continuously, rather than symmetrically projecting or blocking them, is well motivated. The paper also gives credit for reporting standard deviations over three runs and for ablating the two components. However, two load-bearing issues currently prevent acceptance: the algorithm is not uniquely specified, and the theoretical descent-direction claim is not proven.

major comments (2)
  1. [§3.5 / Algorithm 1] Algorithm 1 (lines 3–15) defines a different update from the one described in §3.5. In the pseudocode the inner loop is over tasks with k > i, so the last task's gradient is never modified (g_FG^K = g_K) and the ordering of tasks matters. §3.5 states that each task 'sequentially applies the conflict or cooperation rule with all other tasks j ≠ i', which is symmetric. These two rules are not equivalent. The numbers in Tables 1–4 do not reveal which rule was used, and even the sequential version is ambiguous about whether the numerator in lines 8 and 12 uses the raw g_i or the already-updated g_FG^i. This makes the empirical claim unreproducible as stated. Please correct the pseudocode or the text, specify the exact rule, justify the ordering, and release code so the reported gains can be attributed to a single method.
  2. [§3.6, Eqs. (9)–(10)] The proof does not establish the stated conclusion. Eqs. (9)–(10) only show g_FG^i · g_i ≥ 0. To conclude that g_FG = Σ_i g_FG^i is a descent direction for L_total, one needs Σ_i g_FG^i · g_i + Σ_{i≠j} g_FG^i · g_j ≥ 0. The cross terms are not examined and can be negative. As a logical illustration, take g1=(1,0), g2=(-2,0), g_FG^1=g1, g_FG^2=0; then both per-task inner products are nonnegative but g_FG·(g1+g2) = -1. Therefore the surrogate-descent claim is not established. Provide a proof that controls the cross terms using the actual projection rule, or remove the theoretical claim and present §3.6 as intuition.
minor comments (5)
  1. [§3.4 / Algorithm 1] When cos(θ_ij)=0, Eq. (7) gives g_coop_i = 0.5 g_i because the projection term vanishes. Orthogonal gradients are therefore halved, not 'reinforced'. This magnitude shrinkage is not discussed and may be part of why the method works; please analyze or at least acknowledge it.
  2. [§4.1 / Fig. 2] The experiments are described as using SGD for 100 epochs, but the Figure 2 caption refers to 'the standard Adam optimizer' and 'Adam with PCGrad / FlexiGrad'. Please clarify which optimizer is used in the actual experiments and in the visualization.
  3. [Algorithm 1] The epsilon in the denominator of the projection is never specified or ablated. Since the paper emphasizes that FlexiGrad is parameter-free, please state the value of epsilon and explain why it does not affect the qualitative behavior.
  4. [§4.1] For reproducibility, please report random seeds, exact augmentation details, dataset split conventions, and learning-rate schedules. The statement 'standard augmentations' is too vague for a method whose reported gains are 0.5–1.0%.
  5. [Fig. 5] The accuracy-vs-epoch figure lacks axis labels and a legend entry for 'Vanilla single'. Please improve the figure captions so the visualization can be interpreted independently.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: FlexiGrad's improvements are externally tested, not derived from fitted parameters or self-citations.

full rationale

The paper's central empirical claim is that the FlexiGrad gradient-modulation rule improves multi-granularity accuracy on CUB-200-2011, FGVC-Aircraft, and Stanford Cars. This claim is evaluated against external benchmarks and baselines (PCGrad, FGoN, etc.), and the method is parameter-free: 'FlexiGrad introduces no parameters, requires no modifications to the architecture and integrates seamlessly into any optimiser by replacing the raw gradient with g_FG.' The modulation formulas (Eqs. 5–7) are defined directly on raw task gradients and involve no constants fitted to the reported accuracies. The weighting function w_ij=(1+cos θ_ij)/2 is chosen, ablated against alternatives (Table 3), and not derived from the target results, so it is not a fitted input renamed as a prediction. Self-citations appear in related work and baselines (e.g., FGoN [3] is the authors' prior architecture-level gradient-blocking method), but they are not used to justify FlexiGrad's central claim; the claim is supported by the reported experiments. The theoretical interpretation in §3.6 is logically questionable—the step from 'each corrected gradient remains non-ascending for its own task' to 'their sum g_FG constitutes a principled surrogate descent for L_total' does not follow, as cross-terms with other tasks' gradients are unexamined. However, an invalid proof is not circularity: the empirical results do not depend on that proof, and the proof does not reduce the method to its inputs. Similarly, the inconsistency between Algorithm 1 (which updates only against tasks with k>i) and §3.5 (which says 'all other tasks j≠i') is a reproducibility/correctness problem, not a circularity problem. No step in the derivation chain exhibits self-definitional reasoning, fitted-input-called-prediction, load-bearing self-citation, imported uniqueness, ansatz-smuggling via citation, or renaming of a known result as a new derivation. The honest finding is no significant circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 1 invented entities

The central claim rests on the assumption that pairwise gradient geometry captures hierarchy interactions, on a fixed training protocol, and on an unreported epsilon. There are essentially no free parameters in the method itself, but there are a few implicit choices in the experimental setup.

free parameters (2)
  • epsilon in Algorithm 1 = not reported
    Line 8 and line 12 use an additive epsilon in the denominator ||g_j||^2 + epsilon. Its value is not specified in the paper; the result could be sensitive to it for very small gradient norms, though the paper does not tune it explicitly.
  • training hyperparameters (learning rate, epochs, schedule) = not reported in detail
    All baselines use 'identical optimisation settings', but those settings are not fully specified. The relative ranking could depend on these choices.
assumptions (3)
  • domain assumption Pairwise cosine similarity between gradients at two levels is a sufficient signal to distinguish beneficial from harmful gradient interactions in a hierarchy.
    Invoked in §3.2–§3.4: the whole algorithm is prescribed through cos(θ_ij). No evidence is given that higher-order interactions or task-role asymmetries beyond pairwise geometry matter.
  • domain assumption The shared backbone is trained jointly with K task-specific heads whose losses are summed unweighted (§3.1).
    Equation (1) defines the optimisation target as a plain sum of cross-entropy losses; the claim is that the gradient reshaping preserves this target. If the practical training uses loss weighting, the theory in §3.6 would not directly apply.
  • domain assumption Pre-trained ImageNet features are the right starting point and fine-tuning for 100 epochs with SGD is a representative protocol.
    Used in §4.1 to obtain all reported numbers; the method's superiority is only demonstrated in this regime.
invented entities (1)
  • Hierarchical gradient conflict (as an entity to be measured by cosine similarity)
    purpose: Conceptual framing: describes coarse and fine gradients as being in states of conflict or partial agreement that should be surgically modified.
    It is a framing device; no new physical or computational object is introduced beyond the reweighted gradient.

how reviews work

0 comments
Cite this review

Pith. "Pith review of FlexiGrad: Adaptive Gradient Modulation for Hierarchical Fine-Grained Classification." pith.science (2026). https://pith.science/paper/C6K6H5XX

@misc{pith2026260717563,
  author       = {Pith},
  title        = {Pith review of: FlexiGrad: Adaptive Gradient Modulation for Hierarchical Fine-Grained Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C6K6H5XX}},
  note         = {Machine review of arXiv:2607.17563}
}
read the original abstract

Many fine-grained recognition tasks contain hierarchical labels such as order, family and species. Although this supervision should be beneficial, jointly optimising all levels often leads to unstable training because coarse and fine classifiers impose inconsistent gradients on the shared backbone. This hierarchical gradient conflict prevents the model from learning a coherent coarse-to-fine representation. In this paper, we propose FlexiGrad, a simple and parameter-free method that regulates gradient interactions during backpropagation. FlexiGrad removes only the harmful conflicting component when tasks disagree and reinforces the shared direction when they partially agree through a smooth hierarchy-aware weighting function. This produces stable optimisation and preserves both global structure and fine-grained discriminative cues. FlexiGrad integrates into existing architectures without modification while improves multi-granularity accuracy on CUB-200-2011, FGVC-Aircraft and Stanford Cars. The code will be available at PRIS-CV/FlexiGrad.

Figures

Figures reproduced from arXiv: 2607.17563 by the authors.

Figure 1
Figure 1. Fine-Grained Visual Classification (FGVC) with Hierarchical Labels for Multi￾Granularity Recognition. incompatible ways. Coarse classifiers encourage broad and shape-dominant fea￾tures, while fine classifiers depend on highly localised and subtle attributes [2, 3, 14]. During backpropagation their gradients frequently point in inconsistent directions [30], which creates hierarchical gradient conflict. This conflict … view at source ↗
Figure 2
Figure 2. (a) The 3D landscape of the combined multi-task (coarse & fine Classification) objective. (b) & (c) Contour plots of the individual Coarse Task and Fine Task ob￾jectives. (d) Trajectory using the standard Adam optimizer. The gradient vectors of the two tasks at the end of the trajectory are indicated by blue and orange arrows. (e) Trajectory using Adam with PCGrad. (f) Trajectory using Adam with FlexiGrad. regulates… view at source ↗
Figure 3
Figure 3. Overview of our hierarchy classification model and FlexiGrad method. gi and gj represent the gradients backpropagating from classifier Ci and classifier Cj (i, j ∈ K), which perform classification tasks of coarser granularity level and finer granularity level, respectively. θ represents the angle between gi and gj . 3.1 Hierarchical Optimisation Objective Given an image x with labels {y 1 , y2 , . . . , yK}, the bac… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Visualizations of Vanilla single with PCGrad and Ours at Order, Family, and Species granularities on CUB-200-2011. 5 Further Analysis 5.1 Visualization We visualise how FlexiGrad reshapes hierarchical optimisation through three complementary analyses. Across all visual…
Figure 5
Figure 5. Figure 5: Accuracy performance comparison of Vanilla single, Vanilla single+PCGrad, FGoN, and our method on CUB-200-2011 during the first 20 epochs. Class Index Class Index Class Index Order Family Species R elativ e a c c u r a c y g ain (%) [PITH_FULL_IMAGE:figures/full_fig_p…
Figure 6
Figure 6. Figure 6: Relative accuracy gain versus Vanilla single between Vanilla single with PC￾Grad and our method within Order, Family and Species hierarchy on the CUB-200-2011 dataset. 5.2 Ablation of Core Components [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Species-level accuracy improvement(%) versus Vanilla single on the CUB-200- 2011 dataset across different bird families. As illustrated in [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 3 linked inside Pith

  1. [1]

    International Journal of Data Science and Analytics21(1), 77 (2026) 4

    Abdelsamie, M.M., Azab, S.S., Hefny, H.A.: Deep multi-task learning: a review of concepts, methods, and cross-domain applications. International Journal of Data Science and Analytics21(1), 77 (2026) 4

    2026

  2. [2]

    IEEE Transactions on Image Processing29, 4683–4695 (2020) 2

    Chang, D., Ding, Y., Xie, J., Bhunia, A.K., Li, X., Ma, Z., Wu, M., Guo, J., Song, Y.Z.: The devil is in the channels: Mutual-channel loss for fine-grained image classification. IEEE Transactions on Image Processing29, 4683–4695 (2020) 2

    2020

  3. [3]

    In: CVPR (2021) 2, 4, 8, 9

    Chang, D., Pang, K., Zheng, Y., Ma, Z., Song, Y.Z., Guo, J.: Your” flamingo” is my” bird”: Fine-grained, or not. In: CVPR (2021) 2, 4, 8, 9

    2021

  4. [4]

    In: VCIP (2022) 4

    Chen, J., Chang, D., Xie, J., Du, R., Ma, Z.: Cross-layer feature based multi- granularity visual classification. In: VCIP (2022) 4

    2022

  5. [5]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Chen, J., Zhou, Z., Tong, Y., Chang, D., Luo, Y., Ma, Z.: Seeing as experts do: A knowledge-augmented agent for open-set fine-grained visual understanding. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 41446–41455 (2026) 4

    2026

  6. [6]

    In: CVPR (2009) 8

    Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: Imagenet: A large-scale hierarchical image database. In: CVPR (2009) 8

    2009

  7. [7]

    In: ECCV (2020) 8

    Du, R., Chang, D., Bhunia, A.K., Xie, J., Ma, Z., Song, Y.Z., Guo, J.: Fine-grained visual classification via progressive multi-granularity training of jigsaw patches. In: ECCV (2020) 8

    2020

  8. [8]

    IEEE Transactions on Pattern Analysis and Machine Intelligence44(12), 9521– 9535 (2021) 1

    Du, R., Xie, J., Ma, Z., Chang, D., Song, Y.Z., Guo, J.: Progressive learning of category-consistent multi-granularity features for fine-grained visual classification. IEEE Transactions on Pattern Analysis and Machine Intelligence44(12), 9521– 9535 (2021) 1

    2021

Show all 32 references
  1. [9]

    arXiv preprint arXiv:2504.13608 (2025) 4

    Gao, P., Liang, Y., Song, Y., Yang, Z.: Cross-hierarchical bidirectional consistency learning for fine-grained visual classification. arXiv preprint arXiv:2504.13608 (2025) 4

    2025 arXiv

  2. [10]

    IEEE Transactions on Image Processing35, 3395–3410 (2026) 4

    Gao, Y., Chang, D., Yu, B., Qin, H., Diao, M., Chen, L., Liang, K., Ma, Z.: Toward generalizable forgery detection and reasoning. IEEE Transactions on Image Processing35, 3395–3410 (2026) 4

    2026

  3. [11]

    In: BMVC

    Gao, Y., Lin, W., Xu, J., Xu, W., Chen, P.: Self-supervised adversarial training for robust face forgery detection. In: BMVC. p. 718 (2023) 4

    2023

  4. [12]

    kirchdorfer et al

    Kirchdorfer, L., Sesterhenn, T., Bartelt, C., Stuckenschmidt, H., Schott, L., K¨ ohler, J.M.: Investigating uncertainty weighting for multi-task learning: Insights and an- alytical alternative: L. kirchdorfer et al. International Journal of Computer Vision 134(1), 8 (2026) 4

    2026

  5. [13]

    In: ICCV Workshops (2013) 3, 8

    Krause, J., Stark, M., Deng, J., Fei-Fei, L.: 3d object representations for fine- grained categorization. In: ICCV Workshops (2013) 3, 8

    2013

  6. [14]

    Engineering Applications of Artificial Intelligence163, 113123 (2026) 2

    Kuang, W., Li, Z.: Spatial-aware feature enhancement network for fine-grained visual classification. Engineering Applications of Artificial Intelligence163, 113123 (2026) 2

    2026

  7. [15]

    IEEE Transactions on Vehicular Tech- nology68(5), 4204–4212 (2019) 1

    Li, X., Yu, L., Chang, D., Ma, Z., Cao, J.: Dual cross-entropy loss for small- sample fine-grained vehicle classification. IEEE Transactions on Vehicular Tech- nology68(5), 4204–4212 (2019) 1

    2019

  8. [16]

    NeurIPS (2021) 2, 4

    Liu, B., Liu, X., Jin, X., Stone, P., Liu, Q.: Conflict-averse gradient descent for multi-task learning. NeurIPS (2021) 2, 4

    2021

  9. [17]

    In: AAAI (2025) 4 FlexiGrad for Hierarchical FGVC 15

    Liu, E., Wu, Y.C., Huang, X., Gao, C., Wang, R.J., Xue, K., Qian, C.: Pareto set learning for multi-objective reinforcement learning. In: AAAI (2025) 4 FlexiGrad for Hierarchical FGVC 15

    2025

  10. [18]

    In: AAAI (2025) 4

    Liu, Y., Miao, Y., Xia, L.: Direct routing gradient (drgrad): A personalized infor- mation surgery for multi-task learning (mtl) recommendations. In: AAAI (2025) 4

    2025

  11. [19]

    In: ICCV (2025) 4

    Liu, Y., Yang, L., Wang, Y.: Long-tailed classification with multi-granularity se- mantics. In: ICCV (2025) 4

    2025

  12. [20]

    arXiv preprint arXiv:1306.5151 (2013) 1, 3, 8

    Maji, S., Rahtu, E., Kannala, J., Blaschko, M., Vedaldi, A.: Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151 (2013) 1, 3, 8

    2013 arXiv

  13. [21]

    In: W ACV (2025) 4

    Matsubara, Y., Mendula, M., Levorato, M.: A multi-task supervised compression model for split computing. In: W ACV (2025) 4

    2025

  14. [22]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Qin, H., Chang, D., Gao, Y., Tan, Y., Chen, L., Ma, Z.: Increfa: Breaking the static wall of generative model attribution. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 35405–35415 (2026) 4

    2026

  15. [23]

    arXiv preprint arXiv:2505.15217 (2025) 4

    Qin, H., Chang, D., Gao, Y., Yu, B., Chen, L., Ma, Z.: Multimodal conditional in- formation bottleneck for generalizable ai-generated image detection. arXiv preprint arXiv:2505.15217 (2025) 4

    2025 arXiv

  16. [24]

    In: ICCV (2017) 10

    Selvaraju, R.R., Cogswell, M., Das, A., Vedantam, R., Parikh, D., Batra, D.: Grad- cam: Visual explanations from deep networks via gradient-based localization. In: ICCV (2017) 10

    2017

  17. [25]

    Wah, C., Branson, S., Welinder, P., Perona, P., Belongie, S.: The caltech-ucsd birds-200-2011 dataset (2011) 1, 3, 8

    2011

  18. [26]

    Transactions on Image Processing (2024) 8

    Wang, J., Xu, Q., Jiang, B., Luo, B., Tang, J.: Multi-granularity part sampling attention for fine-grained visual classification. Transactions on Image Processing (2024) 8

    2024

  19. [27]

    Advances in Neural Information Processing Systems38(2026) 4

    Wang, X., Xu, S., Xiangxuan, S., Zhang, Y., Diao, M., Duan, X., Liang, K., Ma, Z., et al.: Cinetechbench: A benchmark for cinematographic technique understanding and generation. Advances in Neural Information Processing Systems38(2026) 4

    2026

  20. [28]

    Wang, X., Zhang, Y., Zhang, X., Yan, H., Diao, M., Xu, S., Yan, Z., Li, H., Liang, K., Ma, Z.: Detailverifybench: A benchmark for dense hallucination localization in long image captions (2026) 4

    2026

  21. [29]

    In: ECCV (2018) 8

    Yang, Z., Luo, T., Wang, D., Hu, Z., Gao, J., Wang, L.: Learning to navigate for fine-grained classification. In: ECCV (2018) 8

    2018

  22. [30]

    NeurIPS (2020) 2, 4, 8

    Yu, T., Kumar, S., Gupta, A., Levine, S., Hausman, K., Finn, C.: Gradient surgery for multi-task learning. NeurIPS (2020) 2, 4, 8

    2020

  23. [31]

    Pattern Recognition (2024) 8

    Zhang, Z.C., Chen, Z.D., Wang, Y., Luo, X., Xu, X.S.: A vision transformer for fine- grained classification by reducing noise and enhancing discriminative information. Pattern Recognition (2024) 8

    2024

  24. [32]

    In: ICCV (2025) 4

    Zheng, H., Yang, S., He, Z., Yang, J., Huang, Z.: Hierarchical cross-modal prompt learning for vision-language models. In: ICCV (2025) 4

    2025

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.