Pith. sign in

REVIEW 4 major objections 7 minor 52 references

Learning Semantic-Aware Threshold for Multi-Label Image Recognition with Partial Labels

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Class-specific thresholds learned from known-sample score distributions improve partial-label multi-label recognition by up to 5.2 average mAP points.

desk verdict Genuine, modest improvement for partial-label multi-label recognition; the gains look real, but the statistical significance claim is under-documented and needs a proper revision before it is fully convincing. read the letter →

arxiv 2507.23263 v1 pith:2OWQY6UI submitted 2025-07-31 cs.CV

Haoxian Ruan , Zhihua Xu , Zhijing Yang , Guang Ma , Jieming Xie , Changxiang Fan , Tianshui Chen This is my paper
classification cs.CV
keywords multi-labelimagerecognitionpartiallabelspseudo-labelingclass-specificthresholdsthresholdlearningdifferentialrankinglossscoredistributionestimationMS-COCO
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

Multi-label recognition models trained on partially annotated data typically generate pseudo-labels for unknown labels by thresholding prediction scores, but a single global threshold ignores the fact that categories have different score distributions and different learning difficulties. The paper claims that this threshold should be learned per category from the score distribution of already-known samples, and it introduces Semantic-Aware Threshold Learning (SATL) to do exactly that. SATL estimates, for every category, quantile boundaries of the known positive and negative score distributions, sets the category threshold at the larger boundary, and updates it smoothly during training; a differential ranking loss then pushes positive scores above and negative scores below the threshold. Plugging SATL into two existing partial-label frameworks improves average mAP from 73.1 to 76.0 on MS-COCO and from 40.1 to 45.3 on VG-200, with consistent gains at known-label proportions from 5% to 50%. If correct, the paper shows that simple distribution-aware threshold estimation is a general drop-in upgrade for pseudo-label-based partial-label multi-label learning.

What carries the argument

The load-bearing object is the per-category threshold $\tau_c$, estimated as $\tau^*_c = \max\{\tau^-_c, \tau^+_c\}$, where $\tau^-_c$ and $\tau^+_c$ are the quantile boundaries of the known negative and positive score distributions defined by $\mathrm{Prob}[\hat{p}^n_c \le \tau^\pm_c] = \kappa^\pm$. This converts threshold selection into a distributional estimation problem: use known labels to build score histograms per category, read off the quantile boundaries, and update the running threshold by $\tau_c(t+1) = \gamma \tau_c(t) + (1-\gamma)\tau^*_c(t)$. The accompanying differential ranking loss uses the signed distance $d^n_c = \max(0, \hat{p}^n_c - \tau_c)$ to encourage positive known labels to sit above the threshold and negative known labels below it, widening the separation that the threshold exploits.

What would settle it

Measure, per category, the divergence between the score distribution of held-out unknown labels and the distribution estimated from known labels (for example with a Kolmogorov-Smirnov statistic) under the 5%-known-label setting; if for many categories the divergence is large and the threshold error tracks it, the central approximation fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that pseudo-label thresholds in MLR-PL should be category-specific, dynamically updated, and derived from the statistical properties of the model's own confidence scores on known labels, rather than fixed or globally decayed. On this view, the model's score distribution for known samples in a category approximates the distribution for unknown samples in the same category, so the quantiles of known positive and known negative distributions can be used to place a threshold that balances precision and recall of pseudo-labels. The proposed SATL framework operationalizes this with a Semantic-Aware Threshold Estimation module and a Differential Ranking Loss, and the experiments show that adding these two components to the SST and HST baselines produces higher mAP, OF1, and CF1 under every tested known-label proportion, with paired t-tests reported at p<0.05 against all compared methods.

Load-bearing premise

The load-bearing premise is that, within each category, the model's prediction scores for samples whose labels are known behave like the scores for samples whose labels are unknown, even though known and unknown sets differ in size and annotation completeness.

Editorial extensions

If this is right

  • Plugging SATL into SST improves average mAP on MS-COCO from 73.1 to 76.0 and on VG-200 from 40.1 to 45.3, with gains at every known-label proportion tested (5%-50%).
  • Plugging SATL into HST similarly improves average mAP from 74.5 to 76.5 on MS-COCO and from 42.6 to 45.3 on VG-200.
  • Ablations show each component contributes: SATE alone raises SST's average mAP to 75.6/44.8 and DRL alone to 75.3/44.2 on MS-COCO/VG-200, with the full method reaching 76.0/45.3.
  • Paired t-tests against all compared methods give p<0.05, so the reported improvements are unlikely to be seed noise under the paper's protocol.
  • Since SATL is designed as a module around existing pseudo-label frameworks, any MLR-PL method that generates confidence scores can adopt class-specific threshold learning the same way.

Reading between the lines

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

  • Beyond the paper, the same quantile-based threshold recipe could be ported to any confidence-score pseudo-labeling pipeline, for example semi-supervised detection or weakly supervised retrieval, since it only needs a per-class score histogram and known labels.
  • The paper's own limitation statement suggests that if annotation is biased rather than random -- for example, only easy examples are labeled -- the known-score distribution will not track the unknown distribution, and thresholds will be miscalibrated; a robust variant might explicitly model label-completeness per category.
  • Because the reported overhead is 2-4 extra hours of training time, a practical extension would estimate thresholds from cached mini-batch score statistics instead of full-dataset histograms, at some cost in statistical stability.
  • The ablation numbers suggest most of the gain comes from SATE alone, with DRL adding roughly half a point; one could test whether an even simpler margin-based regularizer would capture the same separation at lower complexity.
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 / 7 minor

Summary. The paper proposes SATL, a plug-in module for multi-label recognition with partial labels (MLR-PL). SATL estimates category-specific pseudo-label thresholds from the output score distributions of known positive and negative samples, under the assumption that known and unknown samples in the same category have similar score distributions. A second component, DRL, is designed to widen the separation between those distributions. The authors integrate SATL into the SST and HST baselines and report consistent mAP, OF1, and CF1 improvements on MS-COCO and VG-200 across label proportions from 5% to 50%, together with ablations, pseudo-label quality analyses, hyperparameter studies, and a paired t-test.

Significance. If the empirical claims hold, the core idea is simple and practically useful: replacing a global or heuristic pseudo-label threshold with per-category thresholds estimated from score quantiles yields consistent gains on two standard benchmarks, at modest extra training cost. The paper also provides component-level ablations showing that both SATE and DRL contribute, and it reports pseudo-label precision/recall analyses. However, the statistical evidence for the central 'statistically superior' claim is not reproducible as written, and two technical descriptions (Eq. 3 and the DRL gradient behavior) need clarification or correction before the contribution can be fully credited.

major comments (4)
  1. [Section 5.4.3, Table 3, Eq. (17)] The paired t-test is under-specified and the reported p-values cannot be reproduced from the results in Table 1. The paper never states the number of paired samples n, the number of independent random seeds, or the pairing unit. Recomputing from the six label-proportion mAPs in Table 1 with n=6 gives t≈4.93 (p≈0.004) for SST+SATL vs SST and t≈5.03 (p≈0.004) for HST+SATL vs HST, neither matching the reported t=4.8824/p=0.0019 or t=3.9761/p=0.0132. Since the paper uses these tests to conclude statistical superiority, the full paired per-seed data (or at least n and the per-seed mAP vectors) must be provided.
  2. [Section 4.2, Eq. (3)] Equation (3) is notationally ambiguous about which distribution each quantile refers to. It writes Prob[e^p_n_c <= tau^±_c] = kappa^± for n=1,...,N_L, where N_L appears to denote all labeled samples, with no separation into positive and negative samples and no category-specific restriction. As written, tau^+_c and tau^-_c are not defined by the two distributions described in the text. The equation should explicitly state that positive known samples are used to compute tau^+_c with kappa^+ and negative known samples are used to compute tau^-_c with kappa^-, and should index the sample set by the category c.
  3. [Section 4.3, Eqs. (6)-(8)] The DRL as written gives zero gradient to samples on the wrong side of the threshold. For a positive sample with e^p_n_c < tau_c, we have d_n_c = max(0, e^p_n_c - tau_c) = 0 and s_n_c = 1, so the loss term is a constant 1 and the gradient with respect to e^p_n_c is zero. The same happens for negative samples below the threshold. Therefore the statement that 'positive and negative samples ... are propelled in opposite directions' is not an accurate description of the loss. Please clarify the intended behavior or modify the loss so that misclassified samples also receive a gradient.
  4. [Section 3 and Section 5.6] The core assumption that known and unknown output distributions within a category are similar is supported only by the qualitative visualization in Figure 2, and the paper itself in Section 5.6 acknowledges that this alignment can break under a domain gap. Since the entire threshold estimation procedure depends on this assumption, a quantitative per-category comparison (for example, Wasserstein distance or a Kolmogorov-Smirnov statistic between known and unknown score distributions) would materially strengthen the paper's central motivation.
minor comments (7)
  1. [Keywords] The keyword list contains a typo: 'Label Leraning' should be 'Label Learning'.
  2. [Section 5.3 and Tables 1-2] The method is called 'ML-GCN' in the text but 'GCN-ML' in the tables; please unify the naming.
  3. [Section 5.5.4] The sentence 'We choose the precision and recall rates when the model achieves the optimal epoch performance respectively' is unclear; please specify the epoch-selection rule and whether the optimal epoch is chosen per metric or per model.
  4. [Table 4] The ablation results in Table 4 are reported without standard deviations, unlike Table 1; please either report standard deviations or state explicitly that these are single-run results.
  5. [Figure 10] The axis labels are ambiguous ('Values of gamma (%)' with a percentage sign); please specify what quantity is shown on each axis.
  6. [Section 5.5.1 and Figure 7] The phrase 'average precision' in Figure 7 should be defined; the text appears to mean precision averaged over categories, not the standard average-precision metric.
  7. [Figure 9] The annotation 'HST + LSAT' appears to be a typo; it should be 'HST + SATL'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found: SATL is an iterative self-training procedure whose thresholds come from the model's own score distributions, and it is evaluated on external benchmarks.

full rationale

The central derivation chain is not circular. SATE estimates category-specific thresholds from known-sample score distributions (Eqs. 3-5), and those thresholds generate pseudo-labels (Eq. 2) that are combined with known labels (Eq. 12) and trained with Eq. (13). This is a fixed-point/self-training feedback loop, not a logical reduction of the claimed result to its inputs. The key assumption that known and unknown distributions align is stated explicitly in Section 3 and supported by a visualization that uses ground-truth unknown labels; it may be empirically fragile, but it is not an identity. Hyperparameters such as gamma, kappa-plus, and kappa-minus are tuned in Section 5.5.4, and the final numbers are benchmark mAPs on held-out MS-COCO and VG-200 splits, so no fitted parameter is renamed as a prediction. The use of SST and HST as baselines, including prior work by the same authors, is a standard comparison rather than load-bearing self-citation; no uniqueness theorem or unsupported self-citation is invoked to force the method. The under-specified paired t-test in Section 5.4.3 is a statistical reproducibility concern, not circularity: even if the reported p-values cannot be reproduced from the aggregate numbers in Table 1, that does not make the mAP gains equal to the method's inputs by construction.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central empirical claim rests on a set of tuned hyperparameters (kappa_plus, kappa_minus, gamma, lambda) and on a distribution-alignment assumption; no new physical or model entity is introduced.

free parameters (5)
  • kappa_plus = 0.999
    Quantile of the positive-score distribution used to set the high (precision-oriented) threshold; selected after precision-recall analysis on MS-COCO.
  • kappa_minus = 0.1
    Quantile for the negative-score distribution; selected to keep precision high.
  • gamma = 0.3 (20% labels), 0.5 (50% labels)
    Momentum for threshold EMA; chosen from Figure 10 to balance pseudo-label precision and recall.
  • lambda (DRL weight) = 0.01
    Weight of the differential ranking loss; fixed without a reported sensitivity analysis.
  • initial threshold = 1.0
    During the first training stage, no pseudo-labels are used because the threshold starts at the maximum score value.
assumptions (3)
  • domain assumption Known and unknown samples in the same category have similar prediction-score distributions.
    Invoked in Section 3 and Section 4.2 (Eq. 3-4) to justify estimating unknown distributions from known samples.
  • domain assumption Visual features within a category are semantically similar enough to transfer label information from known to unknown samples.
    Stated in Section 1 and Section 4.2 as the basis for the threshold estimation.
  • domain assumption Pseudo-labels selected by high-confidence thresholds improve the MLR-PL model beyond training on known labels alone.
    Adopted from prior pseudo-labeling frameworks in Section 4.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Semantic-Aware Threshold for Multi-Label Image Recognition with Partial Labels." pith.science (2026). https://pith.science/paper/2OWQY6UI

@misc{pith2026250723263,
  author       = {Pith},
  title        = {Pith review of: Learning Semantic-Aware Threshold for Multi-Label Image Recognition with Partial Labels},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2OWQY6UI}},
  note         = {Machine review of arXiv:2507.23263}
}
read the original abstract

Multi-label image recognition with partial labels (MLR-PL) is designed to train models using a mix of known and unknown labels. Traditional methods rely on semantic or feature correlations to create pseudo-labels for unidentified labels using pre-set thresholds. This approach often overlooks the varying score distributions across categories, resulting in inaccurate and incomplete pseudo-labels, thereby affecting performance. In our study, we introduce the Semantic-Aware Threshold Learning (SATL) algorithm. This innovative approach calculates the score distribution for both positive and negative samples within each category and determines category-specific thresholds based on these distributions. These distributions and thresholds are dynamically updated throughout the learning process. Additionally, we implement a differential ranking loss to establish a significant gap between the score distributions of positive and negative samples, enhancing the discrimination of the thresholds. Comprehensive experiments and analysis on large-scale multi-label datasets, such as Microsoft COCO and VG-200, demonstrate that our method significantly improves performance in scenarios with limited labels.

Figures

Figures reproduced from arXiv: 2507.23263 by the authors.

Figure 1
Figure 1. Positive and negative distributions of model’s output scores in four [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Known and unknown distributions of model’s output scores within [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An overall illustration of the proposed learning semantic-aware threshold (SATL) framework. The upper part is the overall pipeline that consists of [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: An illustration of when the positive and negative samples become [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Comparison of training time of baseline (blue) and baseline with SATL module (orange) in different datasets. 7 [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Recall rate in pseudo-labels under the threshold settings of SATL and SST, under the proportion of 20% known labels. Our method recalls more pseudo [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: The average precision of pseudo-labels in unknown labels across all [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Visualization of class-specific thresholds and the distributions of model output scores. The top row represents the distributions of the SATL framework, [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]
Figure 9
Figure 9. Figure 9: A comparative illustration of the recalled labels from HST and our [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Analysis of the precision and recall trend of pseudo-labels under dif [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 13
Figure 13. Figure 13: Analysis of the class loss curves on the MS-COCO dataset with 20% (left) and 50% (right) known labels. SST is adopted as the baseline model in these experiments. distributed in different domains. In such cases, the prediction distributions for known and unknown sample…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 50 canonical work pages

  1. [1]

    Carrillo, D., L ´opez, V .F., Moreno, M.N., 2013. Multi-label classification for recommender systems, in: Trends in Practical Applications of Agents and Multiagent Systems: 11th International Conference on Practical Ap- plications of Agents and Multi-Agent Systems, Springer. pp. 181–188

    work page 2013

  2. [2]

    Chen, T., Lin, J., Yang, Z., Qing, C., Lin, L., 2024a. Learning adaptive spatial coherent correlations for speech-preserving facial expression ma- nipulation, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp. 7267–7276

    work page

  3. [3]

    Knowledge-guided multi-label few-shot learning for general image recognition

    Chen, T., Lin, L., Chen, R., Hui, X., Wu, H., 2022a. Knowledge-guided multi-label few-shot learning for general image recognition. IEEE Trans- actions on Pattern Analysis and Machine Intelligence 44, 1371–1384. doi:10.1109/TPAMI.2020.3025814

    work page arXiv

  4. [4]

    Heterogeneous semantic transfer for multi-label recognition with partial labels

    Chen, T., Pu, T., Liu, L., Shi, Y ., Yang, Z., Lin, L., 2024b. Heterogeneous semantic transfer for multi-label recognition with partial labels. Interna- tional Journal of Computer Vision , 6091–6106

    work page

  5. [5]

    Structured semantic transfer for multi-label recognition with partial labels, in: Proceedings of the AAAI conference on artificial intelligence, pp

    Chen, T., Pu, T., Wu, H., Xie, Y ., Lin, L., 2022b. Structured semantic transfer for multi-label recognition with partial labels, in: Proceedings of the AAAI conference on artificial intelligence, pp. 339–346

    work page

  6. [6]

    Cross-domain facial expression recognition: A unified evaluation benchmark and adver- sarial graph learning

    Chen, T., Pu, T., Wu, H., Xie, Y ., Liu, L., Lin, L., 2021a. Cross-domain facial expression recognition: A unified evaluation benchmark and adver- sarial graph learning. IEEE transactions on pattern analysis and machine intelligence 44, 9887–9903

    work page

  7. [7]

    Dy- namic correlation learning and regularization for multi-label confidence calibration

    Chen, T., Wang, W., Pu, T., Qin, J., Yang, Z., Liu, J., Lin, L., 2024c. Dy- namic correlation learning and regularization for multi-label confidence calibration. IEEE Transactions on Image Processing

    work page

  8. [8]

    Recurrent attentional reinforce- ment learning for multi-label image recognition, in: Proceedings of the AAAI conference on artificial intelligence

    Chen, T., Wang, Z., Li, G., Lin, L., 2018. Recurrent attentional reinforce- ment learning for multi-label image recognition, in: Proceedings of the AAAI conference on artificial intelligence

    work page 2018

Show all 52 references
  1. [9]

    Learning semantic- specific graph representation for multi-label image recognition, in: Pro- ceedings of the IEEE/CVF international conference on computer vision, pp

    Chen, T., Xu, M., Hui, X., Wu, H., Lin, L., 2019a. Learning semantic- specific graph representation for multi-label image recognition, in: Pro- ceedings of the IEEE/CVF international conference on computer vision, pp. 522–531

  2. [10]

    Multi-label image recognition with graph convolutional networks, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Chen, Z.M., Wei, X.S., Wang, P., Guo, Y ., 2019b. Multi-label image recognition with graph convolutional networks, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5177–5186

  3. [11]

    Learning graph con- volutional networks for multi-label recognition and applications

    Chen, Z.M., Wei, X.S., Wang, P., Guo, Y ., 2021b. Learning graph con- volutional networks for multi-label recognition and applications. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 6969– 6983

  4. [12]

    A semantic learning for content-based image retrieval using analytical hierarchy pro- cess

    Cheng, S.C., Chou, T.C., Yang, C.L., Chang, H.Y ., 2005. A semantic learning for content-based image retrieval using analytical hierarchy pro- cess. Expert Systems with Applications 28, 495–505

    2005

  5. [13]

    Ghrs: Graph-based hybrid recom- mendation system with application to movie recommendation

    Darban, Z.Z., Valipour, M.H., 2022. Ghrs: Graph-based hybrid recom- mendation system with application to movie recommendation. Expert Systems with Applications 200, 116850

    2022

  6. [14]

    Learning a deep convnet for multi-label classification with partial labels, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp

    Durand, T., Mehrasa, N., Mori, G., 2019. Learning a deep convnet for multi-label classification with partial labels, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 647–657

    2019

  7. [15]

    Learning to discover multi-class attentional regions for multi-label image recognition

    Gao, B.B., Zhou, H.Y ., 2021. Learning to discover multi-class attentional regions for multi-label image recognition. IEEE Transactions on Image Processing 30, 5920–5932

    2021

  8. [16]

    Interactive multi-label cnn learning with partial labels, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Huynh, D., Elhamifar, E., 2020. Interactive multi-label cnn learning with partial labels, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 9423–9432

    2020

  9. [17]

    Joulin, A., Van Der Maaten, L., Jabri, A., Vasilache, N., 2016. Learning visual features from large weakly supervised data, in: Computer Vision– ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part VII 14, Springer. pp. 67–84

    2016

  10. [18]

    Large loss matters in weakly supervised multi-label classification, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp

    Kim, Y ., Kim, J.M., Akata, Z., Lee, J., 2022. Large loss matters in weakly supervised multi-label classification, in: Proceedings of the IEEE /CVF Conference on Computer Vision and Pattern Recognition, pp. 14156– 14165

    2022

  11. [19]

    Visual genome: Connecting language and vision using crowdsourced dense image anno- tations

    Krishna, R., Zhu, Y ., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalantidis, Y ., Li, L.J., Shamma, D.A., et al., 2017. Visual genome: Connecting language and vision using crowdsourced dense image anno- tations. International journal of computer vision 123, 32–73

    2017

  12. [20]

    Exploiting weakly supervised visual patterns to learn from partial annotations

    Kundu, K., Tighe, J., 2020. Exploiting weakly supervised visual patterns to learn from partial annotations. Advances in Neural Information Pro- cessing Systems 33, 561–572

    2020

  13. [21]

    Instance-aware hashing for multi-label image retrieval

    Lai, H., Yan, P., Shu, X., Wei, Y ., Yan, S., 2016. Instance-aware hashing for multi-label image retrieval. IEEE Transactions on Image Processing 25, 2469–2479

    2016

  14. [22]

    Rethinking pseudo labels for semi-supervised object detection, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp

    Li, H., Wu, Z., Shrivastava, A., Davis, L.S., 2022. Rethinking pseudo labels for semi-supervised object detection, in: Proceedings of the AAAI Conference on Artificial Intelligence, pp. 1314–1322

    2022

  15. [23]

    Technique of image retrieval based on multi-label image annotation, in: 2010 Second Interna- tional Conference on Multimedia and Information Technology, IEEE

    Li, R., Zhang, Y ., Lu, Z., Lu, J., Tian, Y ., 2010. Technique of image retrieval based on multi-label image annotation, in: 2010 Second Interna- tional Conference on Multimedia and Information Technology, IEEE. pp. 10–13

    2010

  16. [24]

    A multi-label, semi-supervised classification approach applied to personality prediction in social media

    Lima, A.C.E., De Castro, L.N., 2014. A multi-label, semi-supervised classification approach applied to personality prediction in social media. Neural Networks 58, 122–130

    2014

  17. [25]

    Microsoft coco: Common objects in context, in: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, Springer

    Lin, T.Y ., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Doll´ar, P., Zitnick, C.L., 2014. Microsoft coco: Common objects in context, in: Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, Spri...

    2014

  18. [26]

    Label-specific feature selection and two-level label recovery for multi-label classification with missing labels

    Ma, J., Chow, T.W., 2019. Label-specific feature selection and two-level label recovery for multi-label classification with missing labels. Neural Networks 118, 110–126

    2019

  19. [27]

    Multi-label zero-shot learning with graph convolutional networks

    Ou, G., Yu, G., Domeniconi, C., Lu, X., Zhang, X., 2020. Multi-label zero-shot learning with graph convolutional networks. Neural Networks 132, 333–341

    2020

  20. [28]

    Dual-perspective semantic-aware representation blending for multi-label image recognition with partial labels

    Pu, T., Chen, T., Wu, H., Shi, Y ., Yang, Z., Lin, L., 2024. Dual-perspective semantic-aware representation blending for multi-label image recognition with partial labels. Expert Systems with Applications , 123526

    2024

  21. [29]

    Au-expression knowledge constrained representation learning for facial expression recognition, in: 2021 IEEE international conference on robotics and automation (ICRA), IEEE

    Pu, T., Chen, T., Xie, Y ., Wu, H., Lin, L., 2021. Au-expression knowledge constrained representation learning for facial expression recognition, in: 2021 IEEE international conference on robotics and automation (ICRA), IEEE. pp. 11154–11161

    2021

  22. [30]

    Asymmetric loss for multi-label classification, in: Proceedings of the IEEE /CVF international conference on computer vision, pp

    Ridnik, T., Ben-Baruch, E., Zamir, N., Noy, A., Friedman, I., Protter, M., Zelnik-Manor, L., 2021. Asymmetric loss for multi-label classification, in: Proceedings of the IEEE /CVF international conference on computer vision, pp. 82–91

    2021

  23. [31]

    In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning

    Rizve, M.N., Duarte, K., Rawat, Y .S., Shah, M., 2021. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning. arXiv preprint arXiv:2101.06329

    2021 arXiv

  24. [32]

    Learning semantic-aware representation in visual-language models for multi-label recognition with partial labels

    Ruan, H., Xu, Z., Yang, Z., Lu, Y ., Qin, J., Chen, T., 2024. Learning semantic-aware representation in visual-language models for multi-label recognition with partial labels. ACM Trans. Multimedia Comput. Com- mun. Appl

    2024

  25. [33]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence

    Sohn, K., Berthelot, D., Carlini, N., Zhang, Z., Zhang, H., Ra ffel, C.A., Cubuk, E.D., Kurakin, A., Li, C.L., 2020. Fixmatch: Simplifying semi- supervised learning with consistency and confidence. Advances in neural information processing systems 33, 596–608

    2020

  26. [34]

    Revisiting unrea- sonable effectiveness of data in deep learning era, in: Proceedings of the IEEE international conference on computer vision, pp

    Sun, C., Shrivastava, A., Singh, S., Gupta, A., 2017. Revisiting unrea- sonable effectiveness of data in deep learning era, in: Proceedings of the IEEE international conference on computer vision, pp. 843–852

    2017

  27. [35]

    Multi-label image categorization with sparse factor representation

    Sun, F., Tang, J., Li, H., Qi, G.J., Huang, T.S., 2014. Multi-label image categorization with sparse factor representation. IEEE Transactions on Image Processing 23, 1028–1037

    2014

  28. [36]

    Image classification with multi-view multi-instance metric learning

    Tang, J., Li, D., Tian, Y ., 2022. Image classification with multi-view multi-instance metric learning. Expert Systems with Applications 189, 116117

    2022

  29. [37]

    Causal multi-label learning for image classification

    Tian, Y ., Bai, K., Yu, X., Zhu, S., 2023. Causal multi-label learning for image classification. Neural Networks 167, 626–637

    2023

  30. [38]

    Multi-label classification: An overview

    Tsoumakas, G., Katakis, I., 2007. Multi-label classification: An overview. International Journal of Data Warehousing and Mining (IJDWM) 3, 1–13

    2007

  31. [39]

    Pico: Contrastive label disambiguation for partial label learning, in: In- ternational Conference on Learning Representations

    Wang, H., Xiao, R., Li, Y ., Feng, L., Niu, G., Chen, G., Zhao, J., 2021. Pico: Contrastive label disambiguation for partial label learning, in: In- ternational Conference on Learning Representations

    2021

  32. [40]

    Cnn- rnn: A unified framework for multi-label image classification, in: Pro- ceedings of the IEEE conference on computer vision and pattern recogni- tion, pp

    Wang, J., Yang, Y ., Mao, J., Huang, Z., Huang, C., Xu, W., 2016. Cnn- rnn: A unified framework for multi-label image classification, in: Pro- ceedings of the IEEE conference on computer vision and pattern recogni- tion, pp. 2285–2294

    2016

  33. [41]

    Image emotion multi-label classification based on multi-graph learning

    Wang, M., Zhao, Y ., Wang, Y ., Xu, T., Sun, Y ., 2023. Image emotion multi-label classification based on multi-graph learning. Expert Systems with Applications 231, 120641. 14

    2023

  34. [42]

    Multi-label image recog- nition by recurrently discovering attentional regions, in: Proceedings of the IEEE international conference on computer vision, pp

    Wang, Z., Chen, T., Li, G., Xu, R., Lin, L., 2017. Multi-label image recog- nition by recurrently discovering attentional regions, in: Proceedings of the IEEE international conference on computer vision, pp. 464–472

    2017

  35. [43]

    Wei, Y ., Xia, W., Lin, M., Huang, J., Ni, B., Dong, J., Zhao, Y ., Yan, S.,

  36. [44]

    Instance- aware representation learning and association for online multi-person tracking

    Wu, H., Hu, Y ., Wang, K., Li, H., Nie, L., Cheng, H., 2019. Instance- aware representation learning and association for online multi-person tracking. Pattern Recognition 94, 25–34

    2019

  37. [45]

    Adahgnn: Adaptive hypergraph neural networks for multi-label image classification, in: Pro- ceedings of the 28th ACM International Conference on Multimedia, pp

    Wu, X., Chen, Q., Li, W., Xiao, Y ., Hu, B., 2020. Adahgnn: Adaptive hypergraph neural networks for multi-label image classification, in: Pro- ceedings of the 28th ACM International Conference on Multimedia, pp. 284–293

    2020

  38. [46]

    Ex- ploit bounding box annotations for multi-label object recognition, in: Pro- ceedings of the IEEE conference on computer vision and pattern recogni- tion, pp

    Yang, H., Tianyi Zhou, J., Zhang, Y ., Gao, B.B., Wu, J., Cai, J., 2016. Ex- ploit bounding box annotations for multi-label object recognition, in: Pro- ceedings of the IEEE conference on computer vision and pattern recogni- tion, pp. 280–288

    2016

  39. [47]

    Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling

    Zhang, B., Wang, Y ., Hou, W., Wu, H., Wang, J., Okumura, M., Shi- nozaki, T., 2021. Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling. Advances in Neural Information Processing Systems 34, 18408–18419

    2021

  40. [48]

    Graph embedding based multi-label zero-shot learning

    Zhang, H., Meng, X., Cao, W., Liu, Y ., Ming, Z., Yang, J., 2023a. Graph embedding based multi-label zero-shot learning. Neural Networks 167, 129–140

  41. [49]

    Zhang, S., Zhang, L., Liu, Z., 2023b. Refined pseudo labeling for source- free domain adaptive object detection, in: ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), IEEE. pp. 1–5

    2023

  42. [50]

    Instance-weighted central similarity for multi- label image retrieval

    Zhang, Z., Peng, H., 2021. Instance-weighted central similarity for multi- label image retrieval. arXiv preprint arXiv:2108.05274

    2021 arXiv

  43. [51]

    Zheng, Y ., Mobasher, B., Burke, R., 2014. Context recommendation using multi-label classification, in: 2014 IEEE /WIC/ACM International Joint Conferences on Web Intelligence (WI) and Intelligent Agent Tech- nologies (IAT), IEEE. pp. 288–295. 15

    2014

  44. [2015]

    IEEE transactions on pattern analysis and machine intelligence 38, 1901– 1907

    Hcp: A flexible cnn framework for multi-label image classification. IEEE transactions on pattern analysis and machine intelligence 38, 1901– 1907

    1901

Pith tools

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