Pith. sign in

REVIEW 4 major objections 5 minor 55 references

Divide, Weight, and Route: Difficulty-Aware Optimization with Dynamic Expert Fusion for Long-tailed Recognition

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read DQRoute claims that adding a class-difficulty signal to reweighting and routing inference through OOD-scored specialized experts lifts few-shot accuracy to 38.6% on CIFAR-100-LT-IR100, versus 33.9% for SADE.

desk verdict Two numbers for the same result and an undefined routing loss make the paper's main claim unverifiable, though the underlying idea is a plausible extension of existing work. read the letter →

arxiv 2508.19630 v1 pith:BA22RLRS submitted 2025-08-27 cs.CV cs.AI

Xiaolei Wei , Yi Ouyang , Haibo Ye This is my paper · ORCID
classification cs.CVcs.AI
keywords long-tailedrecognitionmixtureofexpertsdifficulty-awarereweightingout-of-distributionscoringdynamicroutingclassimbalanceCIFAR-100-LTImageNet-LT
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper sets out to establish that long-tailed recognition should be driven by class difficulty, not just class frequency. DQRoute estimates each class's difficulty from prediction entropy and smoothed accuracy, reweights the loss accordingly, and at inference routes each input through a confidence-weighted combination of three distribution-specialized experts, using out-of-distribution scores as the routing signal. If the claim is right, the rare and intrinsically hard classes that matter most in safety-critical settings stop being the weakest part of the model: on CIFAR-100-LT-IR100, DQRoute reports 38.6% few-shot accuracy against 33.9% for SADE and 23.7% for RIDE, with 51.7% overall accuracy versus 49.4% for SADE. The whole pipeline trains end-to-end, so the difficulty-plus-routing recipe could be added to existing long-tailed systems.

What carries the argument

The machinery is a decentralized routing identity and a difficulty schedule that feed each other. Routing weight α_k(x) = s_k(x) / Σ_j s_j(x) turns each expert's OOD confidence into a soft mixture; the confidence is supervised by the binary label b_k = I[y ∈ C_k] in Eq. 7. The difficulty score d_c = H_c/H_max + λ(1 − A_c/A_max) drives the exponential class weight w_c^(t) in Eq. 2, which is then blended with the frequency prior q_c. Both mechanisms are joined in the total loss L_total = L_cls + λ_OOD L_OOD.

What would settle it

Use the exact CIFAR-100-LT-IR100 setup, take held-out inputs from classes not in the training set, and compare each expert's s_k score with whether that expert is actually correct; the routing hypothesis survives only if s_k ranks the correct expert above the others better than chance. A second check is to recompute Eq. 1's class-wise accuracy strictly on the training set and rerun the full pipeline; if the reported 38.6% few-shot accuracy collapses, the difficulty signal was leaking test statistics.

Watch

Extended reading notes

Core claim

The central claim is that quantity imbalance and intrinsic difficulty are two separate failure modes and that a single joint loop can attack both. DQRoute defines per-class difficulty as a normalized blend of prediction entropy and exponentially smoothed accuracy (Eq. 1), converts that into a class weight that multiplicatively updates each epoch and is blended with a frequency prior (Eqs. 2–3), and uses the resulting weight in the classification loss of every expert. At the same time, each of three experts—trained on all classes, on medium-and-tail classes, and on tail classes only—carries an OOD head whose binary membership label b_k = I[y ∈ C_k] trains a confidence score; these scores are

Load-bearing premise

The load-bearing premise sits in Eq. 7: the score s_k is trained on the binary membership label b_k = I[y ∈ C_k], which only encodes whether the input's class lies in expert k's subset, yet it is used as an OOD confidence measuring distributional alignment—if that transfer fails on novel inputs, the routing weights in Eqs. 5–6 have no grounding, and Eq. 1's accuracy/entropy are never specified as training-measured, leaving test-set leakage possible.

Editorial extensions

If this is right

  • If the results are correct, tail gains do not come at the cost of overall competitiveness: DQRoute's few-shot accuracy is 38.6% while overall accuracy (51.7%) is the best reported on CIFAR-100-LT-IR100 among compared multi-expert methods.
  • The decentralized design means no centralized router is needed at inference; each expert contributes according to its own OOD score, simplifying deployment.
  • The 0.5/0.5 difficulty-quantity blend emerges from the ablation as the operative default; over-weighting either signal degrades both all-class and tail accuracy.
  • The framework is modular and end-to-end trainable; the authors accordingly suggest extending it with alternating-style MoE training to further sharpen specialization.

Reading between the lines

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

  • Reading the ablations, most of the few-shot gain comes from the difficulty reweighting, not the router: OOD-routing alone reached 27.6% tail accuracy, difficulty alone 37.4%, and the combination 38.6%.
  • If s_k truly measures distributional alignment, the same OOD scores could power open-set recognition or selective prediction on novel classes without retraining the router.
  • A deployment version must pin down where Eq. 1's accuracy and entropy are measured; relying on test-set statistics would leak the benchmark, so a validation split or online estimator is the operable choice.
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 DQRoute, a long-tailed recognition framework combining difficulty-aware class reweighting with a mixture-of-experts architecture in which three distribution-specialized experts are fused at inference using per-expert OOD confidence scores. The claimed contributions are: a difficulty measure based on class-wise entropy and accuracy (Eq. 1), a dynamic reweighting scheme (Eqs. 2-3), a decentralized expert routing mechanism (Eqs. 5-8), and joint training (Eqs. 9-10). Experiments report state-of-the-art or competitive accuracy on CIFAR-10/100-LT, ImageNet-LT, and Places-LT, with particular gains on few-shot classes, e.g., 38.6% few-shot accuracy on CIFAR-100-LT IR100 versus 33.9% for SADE. The paper also promises a theoretical analysis of difficulty-gradient interactions, but no such analysis appears in the manuscript.

Significance. If the reported results are reliable, DQRoute would be a useful modular combination of two established ideas---difficulty-aware loss weighting and expert routing---showing that the two components are complementary and that the OOD-based routing improves tail-class performance. The few-shot gains on CIFAR-100-LT are notable. However, the current evidence is internally inconsistent and incomplete: the same setting is reported with two different accuracies, the best-performing loss variant is not defined, and the measurement protocol for the difficulty signal is unspecified. These issues prevent verification of the central claim. The paper provides no code, no error bars, and no per-seed results, so the significance cannot be assessed from the text as it stands.

major comments (4)
  1. [Section 4.2, Tables 1/2 and Table 5] The same configuration, CIFAR-100-LT IR100, is reported as 51.7% overall accuracy in Tables 1 and 2, but as 42.5% in Table 5. The text in §4.2 says DQRoute achieves 42.5% at IR100 in the difficulty-based comparison. No explanation is given for the discrepancy. This is load-bearing: the headline claim of outperforming SADE (49.4%) depends on the 51.7% figure; if 42.5% is the correct value for the full method, the main claim collapses. Please clarify whether Table 5 reports a different variant (e.g., difficulty-only), and if so, state this explicitly and consistently in the text and table headings.
  2. [Section 3.2, Eq. (7), and Table 6] Table 6 identifies EntropyOOD as the best loss for the routing heads (accuracy 51.70%, matching the main result), and the main experiments presumably use it. However, the method section only defines the binary cross-entropy loss in Eq. (7). EntropyOOD is never defined, nor are the other variants (MarginOOD, Focal loss, BCE loss) specified. Without a definition of EntropyOOD, the main results are not reproducible. Please provide the exact loss formulation, including how entropy is computed and how the OOD head is supervised.
  3. [Section 3.1, Eq. (1)] The difficulty score d_c uses 'average entropy' H_c and 'exponentially smoothed accuracy' A_c, but the paper does not state on which data these quantities are measured. If they are computed on the balanced test set, then test labels are used during training, which is a form of leakage. If they are computed on training data, the claim in Section 1 that these are 'test-time signals' and reflect 'realistic inference conditions' is misleading. Please specify the exact measurement protocol (dataset split, timing, batch size, smoothing factor) and justify the choice.
  4. [Section 1, contributions] The first contribution states that the framework is 'supported by a theoretical analysis of difficulty-gradient interactions.' No such analysis appears anywhere in the manuscript, including Section 3 and the appendix-free main text. Either provide the analysis (e.g., a proposition or gradient derivation showing how difficulty weighting affects the optimization) or remove this claim from the contributions.
minor comments (5)
  1. [Tables 1 and 5] The CE baseline for CIFAR-100-LT IR100 is 41.4% in Table 1 but 40.6% in Table 5. Both are labeled as the same setting. Please align these numbers or explain the difference.
  2. [Tables 2 and 3] The method names DBM-BS, DBM-BSL, and DBM-BCL are used across tables without consistency (DBM-BSL appears in Table 2 text, DBM-BCL in Table 3, DBM-BS in Table 2). Please use a single consistent abbreviation and ensure it matches the cited reference.
  3. [Figure 3] The x-axis of Fig. 3 is labeled 'Difficulty Weight Ratio' but the text refers to varying 'weighting ratio' from pure quantity (0) to pure difficulty (1). Clarify whether this is α in Eq. (3) and state the exact grid values used.
  4. [Eq. (3)] The quantity-based weight q_c is not defined. If it is inverse frequency or normalized sample count, state the formula. This matters because Eq. (10) uses ~w_y with y as a function of class, and the reader cannot reproduce the weighting without q_c.
  5. [General] No error bars or standard deviations are reported, and there is no public code. Given the inconsistency in Table 5, please report at least three random seeds for the main results and make the configuration available.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DQRoute's difficulty reweighting and OOD routing are defined from model outputs but are training heuristics, not derivations of an external result; all benchmark claims are externally evaluated.

full rationale

I walked the paper's derivation chain: Eq. (1) defines class difficulty from the model's own entropy and accuracy, Eq. (2) turns it into per-class loss weights, and Eq. (10) applies those weights to the same model's classification loss. This is a self-referential feedback loop, but it is not a case where a claimed 'prediction' or 'result' is equivalent to its input by construction. The difficulty weights are internal training signals, not fitted constants that are then reported as predictions; the final accuracy numbers come from held-out benchmark evaluation against external baselines. Similarly, the OOD routing score in Eq. (7) is trained with a binary label indicating class-subset membership and used to weight experts at inference; this is a supervised auxiliary task and a standard MoE-style gating mechanism, not a circular reduction. There are no load-bearing self-citations: the authors cite prior work by others, and the only mention of ReconBoost is in future work. The promised 'theoretical analysis of difficulty-gradient interactions' is absent, EntropyOOD is never defined, and the CIFAR-100-LT IR100 results are internally inconsistent (51.7% in Table 2 vs. 42.5% in Table 5). These are serious correctness and reproducibility concerns, but they are not circularity: no equation or result reduces to its own inputs in the sense required for a circularity finding. Therefore the circularity score is 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central method rests on several hand-chosen components: the difficulty score definition, the exponential reweighting rule, and the class-group membership score used as a routing confidence. None of these are derived from external constraints, and their controlling hyperparameters are unreported. The 'OOD head' is not an independent entity with outside evidence; it is an internal classifier trained on the same label partition.

free parameters (5)
  • alpha (difficulty vs quantity balance) = not reported; ablation in Fig 3 selects 0.5 on test accuracy
    Eq (3) linearly blends difficulty-based weight with frequency-based weight; chosen by test-set performance in the paper's own ablation.
  • gamma (adaptation sharpness) = not reported
    Eq (2) controls how strongly difficulty changes class weights; no value or sensitivity analysis given.
  • lambda (entropy-accuracy balance in d_c) = not reported
    Eq (1) combines normalized entropy and accuracy into the difficulty score; no value, sweep, or sensitivity reported.
  • lambda_OOD = not reported
    Eq (9) balances classification and OOD losses; no value reported.
  • tau_m and tau_t (expert split thresholds) = not reported
    Section 3.2 defines medium and tail experts using thresholds below tau_m and tau_t; the actual values are absent.
assumptions (5)
  • ad hoc to paper Class difficulty is a linear combination of normalized prediction entropy and normalized accuracy (Definition 1).
    No derivation shows this quantity tracks learnability; it is the paper's core model of difficulty.
  • ad hoc to paper Exponential reweighting in Eq (2) preserves numeric stability and usefully emphasizes difficult classes.
    The update rule is asserted without analysis of convergence or interaction with the loss.
  • ad hoc to paper A binary score trained on bk = I[y in Ck] measures expert relevance for arbitrary inputs.
    Eq (7) labels a sample as in-distribution only if its class lies in the expert's subset; this is a class-group membership classifier, not a true OOD detector.
  • domain assumption Three experts with thresholds tau_m and tau_t span the distributional heterogeneity of the label space.
    The paper assumes fixed head/medium/tail partitions are sufficient for specialization; no adaptive or validated grouping is used.
  • domain assumption Class-wise accuracy and entropy can be estimated online during training without test-set leakage.
    Section 3.1 calls these test-time signals but does not specify whether they are computed on training data, validation data, or test data, or how the smoothing is updated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Divide, Weight, and Route: Difficulty-Aware Optimization with Dynamic Expert Fusion for Long-tailed Recognition." pith.science (2026). https://pith.science/paper/BA22RLRS

@misc{pith2026250819630,
  author       = {Pith},
  title        = {Pith review of: Divide, Weight, and Route: Difficulty-Aware Optimization with Dynamic Expert Fusion for Long-tailed Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BA22RLRS}},
  note         = {Machine review of arXiv:2508.19630}
}
read the original abstract

Long-tailed visual recognition is challenging not only due to class imbalance but also because of varying classification difficulty across categories. Simply reweighting classes by frequency often overlooks those that are intrinsically hard to learn. To address this, we propose \textbf{DQRoute}, a modular framework that combines difficulty-aware optimization with dynamic expert collaboration. DQRoute first estimates class-wise difficulty based on prediction uncertainty and historical performance, and uses this signal to guide training with adaptive loss weighting. On the architectural side, DQRoute employs a mixture-of-experts design, where each expert specializes in a different region of the class distribution. At inference time, expert predictions are weighted by confidence scores derived from expert-specific OOD detectors, enabling input-adaptive routing without the need for a centralized router. All components are trained jointly in an end-to-end manner. Experiments on standard long-tailed benchmarks demonstrate that DQRoute significantly improves performance, particularly on rare and difficult classes, highlighting the benefit of integrating difficulty modeling with decentralized expert routing.

Figures

Figures reproduced from arXiv: 2508.19630 by the authors.

Figure 1
Figure 1. Class-wise accuracy vs. training sample count on CIFAR-100- LT (Imbalance Ratio=100). Accuracy does not monotonically correlate with class size, indicating that quantity-based methods alone are insufficient to ad￾dress long-tailed recognition challenges. To mitigate the effects of class imbalance, most existing approaches focus on modeling class quantity. Techniques such as resampling, loss reweighting, and decouple… view at source ↗
Figure 2
Figure 2. DQRoute Architecture. (1) Input image is processed by a shared fea￾ture extractor. (2) Output featuress are passed to three distribution-specialized experts. (3) Each expert produces both a class prediction, optimized by a classi￾fication loss reweighted by class difficulty and quantity. (4) Each expert produces an OOD score, which is used to fuse expert predictions into the final prediction. the inherent difficulty… view at source ↗
Figure 4
Figure 4. Ablations of different modules. Effect of OOD Experts and Difficulty Weighting. We conduct an ablation study to evaluate the individual and combined effects of the OOD-based expert routing and difficulty-aware reweighting components. As shown in Fig.4, using only OOD-based multiple experts improves overall performance to 49.6%, with strong gains on head (64.5%) and medium (54.7%) classes but limited improve￾ment on … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

55 extracted references · 49 canonical work pages

  1. [1]

    On the effectiveness of out-of-distribution data in self-supervised long-tail learning

    Jianhong Bai, Zuozhu Liu, Hualiang Wang, Jin Hao, Yang Feng, Huanpeng Chu, and Haoji Hu. On the effectiveness of out-of-distribution data in self-supervised long-tail learning. arXiv preprint arXiv:2306.04934 , 2023

    arXiv 2023

  2. [2]

    Eme: Energy-based multiexpert model for long-tailed remote sensing image classification

    Yu Bai, Shuai Shao, Shiyuan Zhao, Weifeng Liu, Dapeng Tao, and Baodi Liu. Eme: Energy-based multiexpert model for long-tailed remote sensing image classification. IEEE Transactions on Geoscience and Remote Sensing , 62:1–12, 2024

    work page 2024

  3. [3]

    Ace: Ally complementary experts for solving long-tailed recognition in one-shot

    Jiarui Cai, Yizhou Wang, and Jenq-Neng Hwang. Ace: Ally complementary experts for solving long-tailed recognition in one-shot. In Proceedings of the IEEE/CVF international conference on computer vision , pages 112–121, 2021

    work page 2021

  4. [4]

    Learning imbalanced datasets with label-distribution-aware margin loss

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Advances in neural information processing systems, 32, 2019

    2019

  5. [5]

    Area: adaptive reweighting via effective area for long-tailed classi- fication

    Xiaohua Chen, Yucan Zhou, Dayan Wu, Chule Yang, Bo Li, Qinghua Hu, and Weiping Wang. Area: adaptive reweighting via effective area for long-tailed classi- fication. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19277–19287, 2023

    work page 2023

  6. [6]

    Remix: rebalanced mixup

    Hsin-Ping Chou, Shih-Chieh Chang, Jia-Yu Pan, Wei Wei, and Da-Cheng Juan. Remix: rebalanced mixup. In European conference on computer vision , pages 95–

    work page

  7. [7]

    Reslt: Resid- ual learning for long-tailed recognition

    Jiequan Cui, Shu Liu, Zhuotao Tian, Zhisheng Zhong, and Jiaya Jia. Reslt: Resid- ual learning for long-tailed recognition. IEEE transactions on pattern analysis and machine intelligence , 45(3):3695–3706, 2022

    work page 2022

  8. [8]

    Parametric con- trastive learning

    Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric con- trastive learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 715–724, 2021

    work page 2021

Show all 55 references
  1. [9]

    Class- balanced loss based on effective number of samples

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class- balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9268– 9277, 2019

    2019

  2. [10]

    Global and local mixture consistency cumulative learning for long-tailed visual recogni- tions

    Fei Du, Peng Yang, Qi Jia, Fengtao Nan, Xiaoting Chen, and Yun Yang. Global and local mixture consistency cumulative learning for long-tailed visual recogni- tions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15814–15823, 2023

    2023

  3. [11]

    Exploring classification equilib- rium in long-tailed object detection

    Chengjian Feng, Yujie Zhong, and Weilin Huang. Exploring classification equilib- rium in long-tailed object detection. In Proceedings of the IEEE/CVF International conference on computer vision , pages 3417–3426, 2021

    2021

  4. [12]

    Shrec’22 track: Open-set 3d object retrieval

    Yifan Feng, Yue Gao, Xibin Zhao, Yandong Guo, Nihar Bagewadi, Nhat-Tan Bui, Hieu Dao, Shankar Gangisetty, Ripeng Guan, Xie Han, et al. Shrec’22 track: Open-set 3d object retrieval. Computers & Graphics , 107:231–240, 2022

    2022

  5. [13]

    Dynamic mixup for multi-label long-tailed food ingredient recognition

    Jixiang Gao, Jingjing Chen, Huazhu Fu, and Yu-Gang Jiang. Dynamic mixup for multi-label long-tailed food ingredient recognition. IEEE Transactions on Multi- media, 25:4764–4773, 2022. Divide, Weight, and Route 13

    2022

  6. [14]

    Long-tailed out-of-distribution detection: Prioritizing attention to tail

    Yina He, Lei Peng, Yongcun Zhang, Juanjuan Weng, Shaozi Li, and Zhiming Luo. Long-tailed out-of-distribution detection: Prioritizing attention to tail. In Proceed- ings of the AAAI Conference on Artificial Intelligence, volume 39, pages 3446–3454, 2025

    2025

  7. [15]

    Disentangling label distribution for long-tailed visual recognition

    Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 6626–6636, 2021

    2021

  8. [16]

    Recon- boost: Boosting can achieve modality reconcilement

    Cong Hua, Qianqian Xu, Shilong Bao, Zhiyong Yang, and Qingming Huang. Recon- boost: Boosting can achieve modality reconcilement. In International Conference on Machine Learning , pages 19573–19597, 2024

    2024

  9. [17]

    Openworldauc: Towards unified evaluation and optimization for open-world prompt tuning

    Cong Hua, Qianqian Xu, Zhiyong Yang, Zitai Wang, Shilong Bao, and Qingming Huang. Openworldauc: Towards unified evaluation and optimization for open-world prompt tuning. In Forty-second International Conference on Machine Learning , 2025

    2025

  10. [18]

    Hierarchical set-to-set represen- tation for 3-d cross-modal retrieval

    Yu Jiang, Cong Hua, Yifan Feng, and Yue Gao. Hierarchical set-to-set represen- tation for 3-d cross-modal retrieval. IEEE TNNLS , pages 1–13, 2023

    2023

  11. [19]

    Decoupling representation and classifier for long- tailed recognition

    Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long- tailed recognition. arXiv preprint arXiv:1910.09217 , 2019

    1910 arXiv

  12. [20]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  13. [21]

    Hybrid generative fusion for efficient and privacy-preserving face recogni- tion dataset generation

    Feiran Li, Qianqian Xu, Shilong Bao, Boyu Han, Zhiyong Yang, and Qingming Huang. Hybrid generative fusion for efficient and privacy-preserving face recogni- tion dataset generation. arXiv preprint arXiv:2508.10672 , 2025

    2025 arXiv

  14. [22]

    One image is worth a thousand words: A usability preservable text-image collaborative erasing framework

    Feiran Li, Qianqian Xu, Shilong Bao, Zhiyong Yang, Xiaochun Cao, and Qingming Huang. One image is worth a thousand words: A usability preservable text-image collaborative erasing framework. In Forty-second International Conference on Ma- chine Learning, 2025

    2025

  15. [23]

    Size-invariance matters: Rethinking metrics and losses for imbalanced multi-object salient object detection

    Feiran Li, Qianqian Xu, Shilong Bao, Zhiyong Yang, Runmin Cong, Xiaochun Cao, and Qingming Huang. Size-invariance matters: Rethinking metrics and losses for imbalanced multi-object salient object detection. In Proceedings of the 41st International Conference on Machine Learnin...

    2024

  16. [24]

    Metasaug: Meta semantic augmentation for long-tailed visual recognition

    Shuang Li, Kaixiong Gong, Chi Harold Liu, Yulin Wang, Feng Qiao, and Xinjing Cheng. Metasaug: Meta semantic augmentation for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, pages 5212–5221, 2021

    2021

  17. [25]

    Focal loss for dense object detection

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ ar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision , pages 2980–2988, 2017

    2017

  18. [26]

    Large-scale long-tailed recognition in an open world

    Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 2537– 2546, 2019

    2019

  19. [27]

    Out-of- distribution detection in long-tailed recognition with calibrated outlier class learn- ing

    Wenjun Miao, Guansong Pang, Xiao Bai, Tianqi Li, and Jin Zheng. Out-of- distribution detection in long-tailed recognition with calibrated outlier class learn- ing. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 38, pages 4216–4224, 2024

    2024

  20. [28]

    Balanced meta- softmax for long-tailed visual recognition

    Jiawei Ren, Cunjun Yu, Xiao Ma, Haiyu Zhao, Shuai Yi, et al. Balanced meta- softmax for long-tailed visual recognition. Advances in neural information process- ing systems , 33:4175–4186, 2020. 14 X. Wei et al

    2020

  21. [29]

    Mol: Joint estimation of micro-expression, optical flow, and land- mark via transformer-graph-style convolution

    Zhiwen Shao, Yifan Cheng, Feiran Li, Yong Zhou, Xuequan Lu, Yuan Xie, and Lizhuang Ma. Mol: Joint estimation of micro-expression, optical flow, and land- mark via transformer-graph-style convolution. IEEE Transactions on Pattern Analysis and Machine Intelligence , pages 1–14, 2025

    2025

  22. [30]

    Identity-invariant representation and transformer-style relation for micro- expression recognition

    Zhiwen Shao, Feiran Li, Yong Zhou, Hao Chen, Hancheng Zhu, and Rui Yao. Identity-invariant representation and transformer-style relation for micro- expression recognition. Applied Intelligence, pages 19860–19871, 2023

    2023

  23. [31]

    Joint facial action unit recognition and self-supervised optical flow estimation

    Zhiwen Shao, Yong Zhou, Feiran Li, Hancheng Zhu, and Bing Liu. Joint facial action unit recognition and self-supervised optical flow estimation. Pattern Recog- nition Letters, pages 70–76, 2024

    2024

  24. [32]

    Difficulty-net: Learning to predict difficulty for long-tailed recognition

    Saptarshi Sinha and Hiroki Ohashi. Difficulty-net: Learning to predict difficulty for long-tailed recognition. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision , pages 6444–6453, 2023

    2023

  25. [33]

    Class-wise difficulty- balanced loss for solving class-imbalance

    Saptarshi Sinha, Hiroki Ohashi, and Katsuyuki Nakamura. Class-wise difficulty- balanced loss for solving class-imbalance. In Proceedings of the Asian conference on computer vision , 2020

    2020

  26. [34]

    Difficulty-aware balancing margin loss for long-tailed recognition

    Minseok Son, Inyong Koo, Jinyoung Park, and Changick Kim. Difficulty-aware balancing margin loss for long-tailed recognition. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 39, pages 20522–20530, 2025

    2025

  27. [35]

    Equalization loss v2: A new gradient balance approach for long-tailed object detection

    Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 1685–1694, 2021

    2021

  28. [36]

    Equalization loss for long-tailed object recognition

    Jingru Tan, Changbao Wang, Buyu Li, Quanquan Li, Wanli Ouyang, Changqing Yin, and Junjie Yan. Equalization loss for long-tailed object recognition. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 11662–11671, 2020

    2020

  29. [37]

    Partial and asymmetric contrastive learning for out-of- distribution detection in long-tailed recognition

    Haotao Wang, Aston Zhang, Yi Zhu, Shuai Zheng, Mu Li, Alex J Smola, and Zhangyang Wang. Partial and asymmetric contrastive learning for out-of- distribution detection in long-tailed recognition. In International Conference on Machine Learning, pages 23446–23458. PMLR, 2022

    2022

  30. [38]

    Seesaw loss for long-tailed instance segmentation

    Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 969...

    2021

  31. [39]

    Contrastive learn- ing based hybrid networks for long-tailed image classification

    Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learn- ing based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 943–952, 2021

    2021

  32. [40]

    The devil is in classification: A simple framework for long-tail instance segmentation

    Tao Wang, Yu Li, Bingyi Kang, Junnan Li, Junhao Liew, Sheng Tang, Steven Hoi, and Jiashi Feng. The devil is in classification: A simple framework for long-tail instance segmentation. In Computer Vision–ECCV 2020: 16th European Confer- ence, Glasgow, UK, August 23–28, 2020, Pro...

    2020

  33. [41]

    Long- tailed recognition by routing diverse distribution-aware experts

    Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella X Yu. Long- tailed recognition by routing diverse distribution-aware experts. arXiv preprint arXiv:2010.01809, 2020

    2010 arXiv

  34. [42]

    Eat: Towards long-tailed out- of-distribution detection

    Tong Wei, Bo-Lin Wang, and Min-Ling Zhang. Eat: Towards long-tailed out- of-distribution detection. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 15787–15795, 2024. Divide, Weight, and Route 15

    2024

  35. [43]

    Adversarial robust- ness under long-tailed distribution

    Tong Wu, Ziwei Liu, Qingqiu Huang, Yu Wang, and Dahua Lin. Adversarial robust- ness under long-tailed distribution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8659–8668, 2021

    2021

  36. [44]

    Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification

    Liuyu Xiang, Guiguang Ding, and Jungong Han. Learning from multiple experts: Self-paced knowledge distillation for long-tailed classification. InComputer Vision– ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Pro- ceedings, Part V 16 , pages 247–263. Spr...

    2020

  37. [45]

    A re-balancing strategy for class-imbalanced classification based on instance diffi- culty

    Sihao Yu, Jiafeng Guo, Ruqing Zhang, Yixing Fan, Zizhen Wang, and Xueqi Cheng. A re-balancing strategy for class-imbalanced classification based on instance diffi- culty. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 70–79, 2022

    2022

  38. [46]

    Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation

    Yuhang Zang, Chen Huang, and Chen Change Loy. Fasa: Feature augmentation and sampling adaptation for long-tailed instance segmentation. In Proceedings of the IEEE/CVF international conference on computer vision , pages 3457–3466, 2021

    2021

  39. [47]

    mixup: Beyond empirical risk minimization

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 , 2017

    2017 arXiv

  40. [48]

    Distribution alignment: A unified framework for long-tail visual recognition

    Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2361– 2370, 2021

    2021

  41. [49]

    Self-supervised ag- gregation of diverse experts for test-agnostic long-tailed recognition

    Yifan Zhang, Bryan Hooi, Lanqing Hong, and Jiashi Feng. Self-supervised ag- gregation of diverse experts for test-agnostic long-tailed recognition. Advances in neural information processing systems , pages 34077–34090, 2022

    2022

  42. [50]

    Ltgc: Long- tail recognition via leveraging llms-driven generated content

    Qihao Zhao, Yalun Dai, Hao Li, Wei Hu, Fan Zhang, and Jun Liu. Ltgc: Long- tail recognition via leveraging llms-driven generated content. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 19510–19520, 2024

    2024

  43. [51]

    Ltrl: Boosting long-tail recognition via reflective learning

    Qihao Zhao, Yalun Dai, Shen Lin, Wei Hu, Fan Zhang, and Jun Liu. Ltrl: Boosting long-tail recognition via reflective learning. In European Conference on Computer Vision, pages 1–18. Springer, 2024

    2024

  44. [52]

    Mdcs: More diverse experts with consistency self-distillation for long-tailed recognition

    Qihao Zhao, Chen Jiang, Wei Hu, Fan Zhang, and Jun Liu. Mdcs: More diverse experts with consistency self-distillation for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 11597– 11608, 2023

    2023

  45. [53]

    Improving calibration for long-tailed recognition

    Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16489–16498, 2021

    2021

  46. [54]

    Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition

    Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. Bbn: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 9719–9728, 2020

    2020

  47. [55]

    Balanced contrastive learning for long-tailed visual recognition

    Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Pro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 6908–6917, 2022

    2022

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.