Pith. sign in

REVIEW 4 major objections 5 minor 48 references

The Demon is in Ambiguity: Revisiting Situation Recognition with Single Positive Multi-Label Learning

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 Verb classification in situation recognition is inherently multi-label — one image can show several valid events — and a new benchmark plus GE-VerbMLP model lift multi-label MAP by over 3 points without sacrificing single-label accuracy.

desk verdict The multi-label reformulation of situation-recognition verbs is a legitimate, well-motivated idea, but the new benchmark's construction—especially the re-annotation step that pulls labels back toward the original single-label distribution—means the headline MAP gain is not yet established. read the letter →

arxiv 2508.21816 v1 pith:AW6IMHRS submitted 2025-08-29 cs.CV cs.AI

Yiming Lin , Yuchen Niu , Shang Wang , Kaizhu Huang , Qiufeng Wang , Xiao-Bo Jin This is my paper · ORCID
classification cs.CVcs.AI
keywords situationrecognitionsinglepositivemulti-labellearningverbclassificationevaluationbenchmarkgraphconvolutionalnetworkadversarialtrainingsemanticambiguity
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 argues that situation recognition's verb classification is not a single-label task: because verb categories overlap semantically, one image can legitimately show several events at once. Standard top-1 and top-5 accuracy therefore ignore valid alternative interpretations, and the paper sets out to measure and exploit this ambiguity. Since full multi-label annotation of the large imSitu training set is too expensive, it reformulates verb classification as single positive multi-label learning (SPMLL), where each image keeps only its original single positive label but the model is trained to predict multiple verbs. To evaluate the reformulation, it constructs a 25,200-image multi-label benchmark on the imSitu test split, and it introduces GE-VerbMLP, a model that combines a graph network over semantic label correlations with adversarial training, improving multi-label mean average precision by more than 3 percentage points while keeping single-label accuracy competitive.

What carries the argument

The work is carried by two pieces of machinery. Formulation-wise, SPMLL (single positive multi-label learning) turns the single-label verb labels into partially-observed multi-label labels, with the empirical risk defined over the one positive label plus implicit negatives. Model-wise, GE-VerbMLP freezes a CLIP image encoder, trains an MLP to get image embeddings, and computes cosine-similarity logits against class-center vectors; a graph convolutional network refines those centers using a 'pseudo' label-correlation graph built from cosine similarities of BERT sentence embeddings of FrameNet verb definitions (sparsified by k-nearest neighbours and smoothed), and adversarial training with PGD

What would settle it

Randomly sample images from the new 25,200-image benchmark and have independent annotators list every verb that applies with no candidate restriction and no distribution-alignment re-annotation; then compare the resulting label sets to the benchmark's. If a substantial share of images contain valid verbs absent from the benchmark's ~20.7-candidate set, or if the re-annotation removed labels that independent annotators keep, the benchmark is not a complete ground truth and the MAP improvements are not measuring what they claim.

Watch

Extended reading notes

Core claim

Verb classification in situation recognition is, the paper argues, a multi-label problem: because verb categories overlap semantically, one image can reasonably be labeled with several verbs at once (a classroom scene may show teaching, lecturing, studying, and writing). Single-label training and evaluation thus ignore valid alternative interpretations. The paper formulates verb classification as single positive multi-label learning (SPMLL) and builds a 25,200-image multi-label benchmark on imSitu via a VLM-LLM pre-annotation pipeline with human verification and re-annotation. It presents GE-VerbMLP, which uses a graph convolutional network over semantic label correlations from FrameNet defi

Load-bearing premise

The new benchmark's ground truth is only as complete as the VLM-LLM candidate list (about 20.7 verbs per image, with 85% coverage measured on a 500-image mini-set) and the re-annotation pass that trims labels to match the original single-label distribution; if valid event labels fall outside that list or get trimmed away, the reported MAP numbers do not measure true multi-label performance.

Editorial extensions

If this is right

  • Multi-label evaluation (MAP) becomes a standard companion to top-1/top-5 accuracy for situation recognition, so a model that ranks the true alternative verbs highly is no longer penalized for not matching the single ground-truth choice.
  • Existing SPMLL losses and model designs can be ported to verb classification: the paper benchmarks nine SPMLL methods and shows several already improve MAP over the single-label cross-entropy baseline, with GE-VerbMLP going furthest.
  • Semantic definitions of the label set can stand in for co-occurrence statistics when multi-label annotations are unavailable, which makes the graph-correlation approach applicable to other fine-grained classification tasks.
  • The 3+ point MAP gain, achieved without sacrificing top-1/top-5 accuracy, implies that single-label training is leaving real performance on the table in situation recognition.

Reading between the lines

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

  • If the benchmark's label completeness is not independently verified on a larger sample, the reported MAP gains could partly reflect the restricted candidate set rather than genuine multi-label understanding; exhaustive annotation of a random subsample would settle this.
  • The same SPMLL framing could be applied to semantic role labeling and grounding, where roles and entities also overlap in meaning across situations.
  • The recipe of 'semantic definitions -> correlation graph -> GCN refinement -> adversarial boundary smoothing' is transferable to other domains with overlapping fine-grained categories, such as video action recognition.
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 argues that verb classification in situation recognition is inherently multi-label, because semantically overlapping verb categories can all validly describe an image. It reformulates the task as Single Positive Multi-Label Learning (SPMLL), constructs a 25,200-image multi-label evaluation benchmark from the imSitu test set using a VLM-LLM pre-annotation pipeline plus human annotation, and proposes GE-VerbMLP, which combines a frozen CLIP encoder with a GCN-refined class-center classifier and adversarial training. On the new benchmark, GE-VerbMLP reports MAP 57.0% versus 53.8% for a standard cross-entropy ClipSitu baseline, a >3% improvement, while roughly matching top-1/top-5 accuracy.

Significance. If the benchmark is valid, this is the first large-scale multi-label evaluation resource for situation recognition and provides a new way to assess whether models capture valid alternative event interpretations. The paper also contributes a reasonable SPMLL formulation and a clearly described method. The qualitative claim of ambiguity is independently supported by the 500-image human mini-benchmark and by FrameNet-based semantic similarity priors, which are strengths. However, the headline quantitative claim rests entirely on the new benchmark's integrity, and the current validation is too weak to establish that integrity. The benchmark may systematically undercount true multi-labels, which would make the reported MAP numbers unreliable as a measure of multi-label performance.

major comments (4)
  1. [§IV-B, §IV-C] The benchmark restricts annotations to a candidate set of ~20.73 verbs per image, generated by a single-positive-label-trained verb classifier plus VLM-LLM. Coverage of human labels is reported as 85% (and 98% after the top-100 stage), but this was measured only on the 500-image mini-benchmark, which was itself used to tune the pipeline. A systematic 15% recall gap is not ruled out. The authors should validate final benchmark recall on a random subset of the 25,200 images with full 504-class annotation, and report per-category and per-image recall, not just an aggregate.
  2. [§IV-C] The re-annotation step 'remov[es] inconsistent labels to align the new annotations with the original label distribution'. The original single-label annotation is the very assumption the paper argues is wrong, so using it as the reference distribution can delete valid alternative verbs that conflict with the original label. This biases the benchmark toward the single-label view and undercounts ambiguity. The authors should quantify how many and which labels were removed, and compare the resulting labels against independent full-504 human annotation on a held-out subset.
  3. [§VI-B, Table III] The headline >3% MAP improvement (53.8% to 57.0%) is computed on the new benchmark. Because the candidate-generation pipeline starts from a single-positive-trained classifier and the same pipeline was tuned on the 500-image mini-benchmark, the MAP values are not independent of the models being compared. At minimum, the authors should report MAP under a setting where true labels are obtained by full 504-class annotation on a random sample, and show that the relative ordering of methods is stable. Without this, the central empirical claim is not yet established.
  4. [§V.A.3, Eq. (9)] The text says the class center c_j is randomly initialized, but Eqs. (4), (5), and (9) use c_i = BERT(s_i) as the initial class center. This is internally inconsistent. If C1 is random, the semantic-similarity prior does not enter the model as claimed; if C1 uses BERT embeddings, the text should say so. This is load-bearing for the method's explanation and should be corrected.
minor comments (5)
  1. [Abstract and §I] Typo: 'futher' should be 'further'; 'the formulate, solving, and evaluating' is ungrammatical.
  2. [§IV-D, Table I] The text says 119,272 total annotations, Table I says 119,372. Also Table I reports max labels per image as 21 while Fig. 7 shows a label-count distribution with max 13. These numerical inconsistencies need correction.
  3. [§V.B] Section title 'Incoporating Adversarial Training' should be 'Incorporating Adversarial Training'.
  4. [§VII] Conclusion says 'single forward multi-label learning (SPMLL)'; should be 'single positive multi-label learning'.
  5. [§II.A] Reference formatting: 'Pratt et al. .' has an extra period before the citation; please clean up citation typos throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the central derivation is independent, with benchmark validity as a separate concern.

full rationale

The paper's central claims—that verb classification is inherently multi-label and that GE-VerbMLP improves MAP—are grounded in external evidence rather than in self-referential equations. The ambiguity claim rests on a 500-image mini-benchmark where human annotators could select from all 504 classes (Sec IV-A), and on t-SNE visualizations of CLIP embeddings. The method uses FrameNet semantic similarities (Sec V-A1) as an external prior for label correlations, and adversarial training via standard FGSM/PGD (Sec V-B), sourced from references [3,4] rather than from the authors' own survey [2]. The MAP evaluation is performed on a human-verified multi-label benchmark; although the benchmark's candidate set is restricted by a VLM-LLM pipeline and a re-annotation step aligns labels with the original single-label distribution (Sec IV-C), these are benchmark-validity concerns, not circular reductions. No equation in the paper equates the model's outputs to the benchmark labels, and the human annotations are not generated by the model being evaluated. The paper's self-citation ([2]) is not load-bearing because the adversarial methods are cited to their original sources. The reported 85% coverage of the VLM-LLM candidate set on the mini-benchmark is a missing-support point for the final benchmark's completeness, but that affects correctness and generalizability, not circularity. Therefore, the paper does not exhibit the specific reduction required to claim circularity.

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

The paper's central empirical claims rest on a relatively small set of hand-set hyperparameters and on assumptions about benchmark label completeness and the semantic-similarity proxy. The most consequential are the benchmark coverage and re-annotation assumptions, because they control the validity of every MAP number reported.

free parameters (5)
  • temperature tau = 10
    Eq. 10 divides cosine logits by tau; set empirically to 10 and not derived.
  • graph smoothing hyperparameter s = not stated
    Eq. 6 controls how much weight is assigned to neighbors vs self; value not reported.
  • K nearest neighbors for label graph = small value (best low K per Fig. 10)
    Eq. 5 sparsifies the correlation matrix; K is selected by MAP on the benchmark.
  • GCN depth J = 2 (best per Fig. 10)
    Receptive field depth; MAP drops beyond two layers, so J is tuned on the benchmark.
  • Adversarial training epsilon / PGD iterations = not reported
    FGSM/PGD in Eq. 11-13 need epsilon and step sizes; values are absent, so attacks cannot be replicated.
assumptions (5)
  • domain assumption An image can legitimately contain multiple visual events, so a single positive label is one sample from a set of valid labels.
    Core premise of Sec III and the SPMLL reformulation; supported only by the paper's own mini-benchmark.
  • domain assumption Semantic similarity of FrameNet definitions is a valid proxy for true label co-occurrence.
    Used in Sec V-A1 (Eq. 4-6) to construct the pseudo label-correlation graph that drives GCN.
  • domain assumption VLM-LLM candidate generation covers nearly all true labels on the full 25,200-image benchmark.
    Coverage (98%, then 85%) is measured only on the 500-image mini-benchmark in Sec IV-B but extrapolated to the full benchmark.
  • domain assumption Re-annotation that removes labels to align with the original distribution does not discard valid labels.
    Sec IV-C describes removing inconsistent labels to match the original label distribution, which can suppress legitimate alternatives.
  • standard math GCN message passing, cosine similarity, BCE/focal losses, and FGSM/PGD provide valid optimization machinery.
    Standard tools assumed without proof.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The Demon is in Ambiguity: Revisiting Situation Recognition with Single Positive Multi-Label Learning." pith.science (2026). https://pith.science/paper/AW6IMHRS

@misc{pith2026250821816,
  author       = {Pith},
  title        = {Pith review of: The Demon is in Ambiguity: Revisiting Situation Recognition with Single Positive Multi-Label Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AW6IMHRS}},
  note         = {Machine review of arXiv:2508.21816}
}
read the original abstract

Context recognition (SR) is a fundamental task in computer vision that aims to extract structured semantic summaries from images by identifying key events and their associated entities. Specifically, given an input image, the model must first classify the main visual events (verb classification), then identify the participating entities and their semantic roles (semantic role labeling), and finally localize these entities in the image (semantic role localization). Existing methods treat verb classification as a single-label problem, but we show through a comprehensive analysis that this formulation fails to address the inherent ambiguity in visual event recognition, as multiple verb categories may reasonably describe the same image. This paper makes three key contributions: First, we reveal through empirical analysis that verb classification is inherently a multi-label problem due to the ubiquitous semantic overlap between verb categories. Second, given the impracticality of fully annotating large-scale datasets with multiple labels, we propose to reformulate verb classification as a single positive multi-label learning (SPMLL) problem - a novel perspective in SR research. Third, we design a comprehensive multi-label evaluation benchmark for SR that is carefully designed to fairly evaluate model performance in a multi-label setting. To address the challenges of SPMLL, we futher develop the Graph Enhanced Verb Multilayer Perceptron (GE-VerbMLP), which combines graph neural networks to capture label correlations and adversarial training to optimize decision boundaries. Extensive experiments on real-world datasets show that our approach achieves more than 3\% MAP improvement while remaining competitive on traditional top-1 and top-5 accuracy metrics.

Figures

Figures reproduced from arXiv: 2508.21816 by the authors.

Figure 1
Figure 1. Illustration of Situation Recognition. Given an image, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Illustration of sample embeddings with T-SNE visualization. Different colors represent samples of different categories. [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Three main steps for data collection: 1) We gather [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (6 more)
Figure 6
Figure 6. Figure 6: Two-stage annotation pipeline. (Top) VLM-based im [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 5
Figure 5. Figure 5: According to human annotation, we calculate the [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Top-20 most common labels (left) and per-image label [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: We use semantic relations to regularize the relations [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: T-SNE visualization of different class center vectors. [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: , we investigate the impact of adjusting the nearest neighbor number K while constructing class correlation graph Aˆ. Increasing the amount of K results in a denser graph connection, leading to a more uniform distribution of class center vectors. It can be observed th…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 37 canonical work pages

  1. [1]

    Multi-Label Learning from Single Positive Labels

    E. Cole, O. M. Aodha, T. Lorieul, P. Perona, D. Morris, and N. Jojic, “Multi-label learning from single positive labels,” 2021. [Online]. Available: https://arxiv.org/abs/2106.09708

    work page Pith review arXiv 2021

  2. [2]

    A survey of robust adversarial training in pattern recognition: Fundamental, theory, and methodologies,

    Z. Qian, K. Huang, Q.-F. Wang, and X.-Y . Zhang, “A survey of robust adversarial training in pattern recognition: Fundamental, theory, and methodologies,” Pattern Recognition, vol. 131, p. 108889, 2022

    work page 2022

  3. [3]

    Explaining and harnessing adversarial examples,

    I. J. Goodfellow, J. Shlens, and C. Szegedy, “Explaining and harnessing adversarial examples,” arXiv preprint arXiv:1412.6572 , 2014

    arXiv 2014

  4. [4]

    Towards deep learning models resistant to adversarial attacks,

    A. Madry, A. Makelov, L. Schmidt, D. Tsipras, and A. Vladu, “Towards deep learning models resistant to adversarial attacks,” arXiv preprint arXiv:1706.06083, 2017

    arXiv 2017

  5. [5]

    Generative adversarial nets,

    I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative adversarial nets,” Advances in Neural Information Processing Systems , vol. 27, 2014

    2014

  6. [6]

    Gandef: A gan based adversarial training defense for neural network classifier,

    G. Liu, I. Khalil, and A. Khreishah, “Gandef: A gan based adversarial training defense for neural network classifier,” in ICT Systems Security and Privacy Protection: 34th IFIP TC 11 International Conference, SEC 2019, Lisbon, Portugal, June 25-27, 2019, Proceedings 34 . Springer, 2019, pp. 19–32

    work page 2019

  7. [7]

    Denoising diffusion probabilistic models,

    J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in Neural Information Processing Systems , vol. 33, pp. 6840– 6851, 2020

    2020

  8. [8]

    Weakly supervised multi-label learning via label enhancement

    J. Lv, N. Xu, R. Zheng, and X. Geng, “Weakly supervised multi-label learning via label enhancement.” in IJCAI, 2019, pp. 3101–3107

    work page 2019

Show all 48 references
  1. [9]

    Multi-label learning from single positive labels,

    E. Cole, O. Mac Aodha, T. Lorieul, P. Perona, D. Morris, and N. Jojic, “Multi-label learning from single positive labels,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 933–942

    2021

  2. [10]

    When does label smoothing help?

    R. M ¨uller, S. Kornblith, and G. E. Hinton, “When does label smoothing help?” Advances in Neural Information Processing Systems , vol. 32, 2019

    2019

  3. [11]

    Simple and robust loss design for multi-label learning with missing labels,

    Y . Zhang, Y . Cheng, X. Huang, F. Wen, R. Feng, Y . Li, and Y . Guo, “Simple and robust loss design for multi-label learning with missing labels,” arXiv preprint arXiv:2112.07368 , 2021

    2021 arXiv

  4. [12]

    Large loss matters in weakly supervised multi-label classification,

    Y . Kim, J. M. Kim, Z. Akata, and J. Lee, “Large loss matters in weakly supervised multi-label classification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 14 156–14 165

    2022

  5. [13]

    Bridging the gap between model explanations in partially annotated multi-label classification,

    Y . Kim, J. M. Kim, J. Jeong, C. Schmid, Z. Akata, and J. Lee, “Bridging the gap between model explanations in partially annotated multi-label classification,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 3408–3417

    2023

  6. [14]

    Exploring structured semantic prior for multi label recognition with incomplete labels,

    Z. Ding, A. Wang, H. Chen, Q. Zhang, P. Liu, Y . Bao, W. Yan, and J. Han, “Exploring structured semantic prior for multi label recognition with incomplete labels,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2023, pp. 3398–3407

    2023

  7. [15]

    Revisiting pseudo-label for single- positive multi-label learning,

    B. Liu, N. Xu, J. Lv, and X. Geng, “Revisiting pseudo-label for single- positive multi-label learning,” in International Conference on Machine Learning. PMLR, 2023, pp. 22 249–22 265

    2023

  8. [16]

    Hierarchical prompt learning using clip for multi-label classification with single positive labels,

    A. Wang, H. Chen, Z. Lin, Z. Ding, P. Liu, Y . Bao, W. Yan, and G. Ding, “Hierarchical prompt learning using clip for multi-label classification with single positive labels,” in Proceedings of the 31st ACM Interna- tional Conference on Multimedia , 2023, pp. 5594–5604

    2023

  9. [17]

    Clipsitu: Effectively leveraging clip for conditional predictions in situation recognition,

    D. Roy, D. Verma, and B. Fernando, “Clipsitu: Effectively leveraging clip for conditional predictions in situation recognition,” in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2024, pp. 444–453

    2024

  10. [18]

    Clip-event: Connecting text and images with event structures,

    M. Li, R. Xu, S. Wang, L. Zhou, X. Lin, C. Zhu, M. Zeng, H. Ji, and S.-F. Chang, “Clip-event: Connecting text and images with event structures,” in 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2022, pp. 16 399–16 408

    2022

  11. [19]

    Recurrent models for situation recognition,

    A. Mallya and S. Lazebnik, “Recurrent models for situation recognition,” in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 455–463

    2017

  12. [20]

    Situation recognition with graph neural networks,

    R. Li, M. Tapaswi, R. Liao, J. Jia, R. Urtasun, and S. Fidler, “Situation recognition with graph neural networks,” in Proceedings of the IEEE International Conference on Computer Vision , 2017, pp. 4173–4182

    2017

  13. [21]

    Mixture-kernel graph attention network for situation recognition,

    M. Suhail and L. Sigal, “Mixture-kernel graph attention network for situation recognition,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2019, pp. 10 363–10 372

    2019

  14. [22]

    Attention-based context aware reasoning for situation recognition,

    T. Cooray, N.-M. Cheung, and W. Lu, “Attention-based context aware reasoning for situation recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 4736–4745

    2020

  15. [23]

    Collaborative transformers for grounded situation recognition,

    J. Cho, Y . Yoon, and S. Kwak, “Collaborative transformers for grounded situation recognition,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 19 659–19 668

    2022

  16. [24]

    Ambiguous images with human judgments for robust visual event classification,

    K. Sanders, R. Kriz, A. Liu, and B. Van Durme, “Ambiguous images with human judgments for robust visual event classification,” Advances in Neural Information Processing Systems , vol. 35, pp. 2637–2650, 2022

    2022

  17. [25]

    Situation recognition: Visual semantic role labeling for image understanding,

    M. Yatskar, L. Zettlemoyer, and A. Farhadi, “Situation recognition: Visual semantic role labeling for image understanding,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2016, pp. 5534–5542

    2016

  18. [26]

    Grounded situation recognition,

    S. Pratt, M. Yatskar, L. Weihs, A. Farhadi, and A. Kembhavi, “Grounded situation recognition,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IV

    2020

  19. [27]

    Springer, 2020, pp. 314–332

    2020

  20. [28]

    Grounded situation recognition with transformers,

    J. Cho, Y . Yoon, H. Lee, and S. Kwak, “Grounded situation recognition with transformers,” arXiv preprint arXiv:2111.10135 , 2021

    2021 arXiv

  21. [29]

    Visualizing data using t-sne

    L. Van der Maaten and G. Hinton, “Visualizing data using t-sne.” Journal of Machine Learning Research , vol. 9, no. 11, 2008

    2008

  22. [30]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” arXiv preprint arXiv:1609.02907 , 2016

    2016 arXiv

  23. [31]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International Conference on Machine Learning . PMLR, 2021, pp. 8748–8763

    2021

  24. [32]

    The framenet database and software tools

    C. J. Fillmore, C. F. Baker, and H. Sato, “The framenet database and software tools.” in LREC. Citeseer, 2002

    2002

  25. [33]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” arXiv preprint arXiv:1908.10084 , 2019

    1908 arXiv

  26. [34]

    Focal loss for dense object detection,

    T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss for dense object detection,” in Proceedings of the IEEE International Conference on Computer Vision , 2017, pp. 2980–2988

    2017

  27. [35]

    Learning deep features for discriminative localization,

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Torralba, “Learning deep features for discriminative localization,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2016, pp. 2921–2929

    2016

  28. [36]

    The cityscapes dataset for semantic urban scene understanding,

    M. Cordts, M. Omran, S. Ramos, T. Rehfeld, M. Enzweiler, R. Be- nenson, U. Franke, S. Roth, and B. Schiele, “The cityscapes dataset for semantic urban scene understanding,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , 2016, pp. 3213–3223

    2016

  29. [37]

    Activitynet: A large-scale video benchmark for human activity under- standing,

    F. Caba Heilbron, V . Escorcia, B. Ghanem, and J. Carlos Niebles, “Activitynet: A large-scale video benchmark for human activity under- standing,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 961–970

    2015

  30. [38]

    A survey on deep learning-driven remote sensing image scene understanding: Scene classification, scene retrieval and scene-guided object detection,

    Y . Gu, Y . Wang, and Y . Li, “A survey on deep learning-driven remote sensing image scene understanding: Scene classification, scene retrieval and scene-guided object detection,” Applied Sciences, vol. 9, no. 10, p. 2110, 2019

    2019

  31. [39]

    Learning and understanding dynamic scene activity: a review,

    H. Buxton, “Learning and understanding dynamic scene activity: a review,”Image and Vision Computing, vol. 21, no. 1, pp. 125–136, 2003

    2003

  32. [40]

    Acknowledging the unknown for multi-label learning with single positive labels,

    D. Zhou, P. Chen, Q. Wang, G. Chen, and P.-A. Heng, “Acknowledging the unknown for multi-label learning with single positive labels,” in European Conference on Computer Vision . Springer, 2022, pp. 423– 440

    2022

  33. [41]

    One positive label is sufficient: Single-positive multi-label learning with label enhance- ment,

    N. Xu, C. Qiao, J. Lv, X. Geng, and M.-L. Zhang, “One positive label is sufficient: Single-positive multi-label learning with label enhance- ment,” Advances in Neural Information Processing Systems , vol. 35, pp. 21 765–21 776, 2022

    2022

  34. [42]

    Qwen2.5-vl technical report,

    S. Bai and K. C. et al., “Qwen2.5-vl technical report,” 2025. [Online]. Available: https://arxiv.org/abs/2502.13923

    2025 arXiv

  35. [43]

    Deepseek-vl2: Mixture-of-experts vision- language models for advanced multimodal understanding,

    Z. Wu and X. C. et al., “Deepseek-vl2: Mixture-of-experts vision- language models for advanced multimodal understanding,” 2024. [Online]. Available: https://arxiv.org/abs/2412.10302

    2024 arXiv

  36. [44]

    Deepseek-v3 technical report,

    DeepSeek-AI, A. Liu, and B. F. et al., “Deepseek-v3 technical report,”

  37. [45]

    Co-pseudo labeling and active selection for fundus single-positive multi-label learning,

    T. Hu, W. Zhang, J. Guo, and H. Li, “Co-pseudo labeling and active selection for fundus single-positive multi-label learning,” IEEE Trans- actions on Medical Imaging , 2025

    2025

  38. [46]

    Semantic-guided representation learning for multi-label recognition,

    R. Zhang, H. Qiao, P. Xu, M. Shang, and L. Chen, “Semantic-guided representation learning for multi-label recognition,” arXiv preprint arXiv:2504.03801, 2025

    2025 arXiv

  39. [48]

    Splicemix: A cross-scale and semantic blending augmentation strategy for multi-label image classification,

    L. Wang, Y . Zhan, L. Ma, D. Tao, L. Ding, and C. Gong, “Splicemix: A cross-scale and semantic blending augmentation strategy for multi-label image classification,” IEEE Transactions on Multimedia , 2025

    2025

  40. [2025]

    Available: https://arxiv.org/abs/2412.19437

    [Online]. Available: https://arxiv.org/abs/2412.19437

    arXiv

Pith tools

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