Pith. sign in

REVIEW 3 major objections 5 minor 55 references

Text-guided Weakly Supervised Framework for Dynamic Facial Expression Recognition

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Text-guided visual prompts plus a two-scale temporal network let a weakly supervised video model recognize facial expressions at state-of-the-art accuracy while identifying which frames carry the emotion.

desk verdict The paper's headline accuracy comes from a configuration that omits its claimed key ingredient, and its own ablation shows that ingredient hurts performance. read the letter →

arxiv 2511.10958 v1 pith:BF2I5RX2 submitted 2025-11-14 cs.CV cs.AI

Gunho Jung , Heejo Kong , Seong-Whan Lee This is my paper
classification cs.CVcs.AI
keywords dynamicfacialexpressionrecognitionweaklysupervisedlearningmultipleinstancevision-languagepretrainingvisualpromptingtemporalmodelingvideoemotionframe-levelrelevance
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 tries to establish that the many-to-one labeling problem in dynamic facial expression recognition can be tamed by injecting fine-grained textual descriptions of facial movements into a weakly supervised video classifier. The proposed framework treats each video as a bag of frames, uses a frozen vision-language model to generate text-based emotion descriptions, and learns visual prompts that align those descriptions with each frame's visual features, producing frame-level importance scores. It couples this with a two-scale temporal network that captures both short facial dynamics and long-range emotional flow. The authors report new state-of-the-art results on two benchmarks and show that frames the model deems most influential indeed carry the target emotion. A sympathetic reader would care because this suggests that cheap clip-level labels, plus language-derived semantics, can partially substitute for expensive frame-level annotation.

What carries the argument

The central mechanism is the visual prompt module — a video-conditioned, text-anchored attention process. Given fine-grained text embeddings and frame-level visual features, it computes instance-level alignment scores via cosine similarity with a temperature scale, normalizes them, and forms a visual prompt as their weighted sum over frames. The prompt is added elementwise to the text embedding, passed through an MLP, and combined with coarse text embeddings to produce an enhanced label feature. This feature is used in a softmax over cosine similarities to make the bag prediction under a cross-entropy MIL loss. The module re-weights frames by their semantic relevance to the text-described ex

What would settle it

A concrete check: rerun the DFEW experiments with the fine-grained text descriptions replaced by empty or nonsense phrases (e.g., 'a person') while keeping all other components and hyperparameters identical. If UAR/WAR do not fall substantially below the reported 60.17/71.62, then the gains do not come from text semantics. A second check: take the learned alignment scores and compare the top-ranked frames against human judgments of which frames display the target emotion; if agreement is near chance, the interpretability claim fails.

Watch

Extended reading notes

Core claim

The central claim is that text-guided visual prompting resolves a core weakness of MIL-based dynamic facial expression recognition: the visual diversity and temporal complexity that scatter emotionally relevant frames in feature space. Concretely, the framework constructs an enhanced fine-grained label feature by embedding detailed textual descriptions with a frozen text encoder, computing cosine-similarity alignment scores between those text embeddings and each frame's visual features, distilling a video-conditional visual prompt as a weighted sum of frame features, and fusing this prompt with the text embedding and a coarse class embedding through residual addition and an MLP. Matching thi

Load-bearing premise

The load-bearing premise is that fine-grained textual descriptions such as 'a smiling mouth, widened eyes' are discriminative for all seven emotions and remain valid across different video datasets; if those descriptions are ambiguous, overlapping, or untransferable, the visual prompt alignment will inject noise rather than semantic guidance.

Editorial extensions

If this is right

  • If the central claim holds, weakly supervised video emotion recognition can reach competitive accuracy without frame-level labels, lowering annotation costs.
  • The learned instance-level alignment scores give a free frame-importance map, useful for explaining model decisions and for downstream tasks like keyframe selection.
  • The approach suggests that frozen vision-language encoders plus lightweight trainable temporal and prompt modules are a viable recipe for video recognition under scarce supervision.
  • The ablation showing shallow temporal depths work best implies emotional cues in current benchmarks are localized; models should prioritize short decisive moments over long context.
  • The two-dataset results indicate that fine-grained text guidance transfers across in-the-wild distributions, pointing toward more generalizable emotion models.

Reading between the lines

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

  • Because the fine-grained text descriptions are inherited from prior work and the paper gives only a single example, the framework's success may hinge on those descriptors; a controlled test with automatically generated or per-class optimized text would reveal whether the guidance is transferable or dataset-specific.
  • The 'Highest vs Lowest influence' inference experiment suggests the enhanced label feature could be repurposed as a weakly supervised frame localizer, a natural extension the authors do not develop.
  • The shallow-optimal finding hints that longer, more gradual emotional transitions are underrepresented in current benchmarks; as datasets add longer videos, the multi-grained temporal design may need deeper coarse branches.
  • The visual-prompt mechanism is generic enough to be applied to other weakly supervised video tasks such as action localization or anomaly detection where text semantics are available, but the paper does not test this.
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

3 major / 5 minor

Summary. The paper proposes TG-DFER, a weakly supervised framework for dynamic facial expression recognition (DFER) that combines a CLIP-based vision-language backbone, a multi-grained temporal transformer, and a visual-prompt module that aligns fine-grained textual emotion descriptions with frame-level visual features. Videos are treated as bags of frames with clip-level labels, and the model computes frame relevance through text–visual similarity. Experiments on DFEW and FERV39k report state-of-the-art UAR/WAR (60.17/71.62 on DFEW; 41.50/51.67 on FERV39k), with ablations on temporal depths, prompt configurations, and visual-prompt placement.

Significance. If the reported gains are robust, TG-DFER would make a modest but useful contribution to weakly supervised DFER by showing that text-guided visual prompting can improve MIL-based recognition without frame-level annotations. The paper is clearly written and includes standard comparisons, a fairly thorough ablation study, and qualitative evidence that the learned fine-grained label feature identifies emotionally salient frames. However, the central narrative—that fine-grained textual descriptions such as 'a smiling mouth, widened eyes' are the source of improvement—is not supported by the ablation table that corresponds to the deployed configuration. The absolute gains over DFER-CLIP are sub-percent (0.56 UAR on DFEW, 0.23 UAR on FERV39k), and no error bars or per-fold results are provided, so the state-of-the-art claim is not firmly established.

major comments (3)
  1. [§4.3, Table 5] The reported SOTA configuration (Table 1) corresponds to Table 5, setting h: '[Learnable] Class Add' (UAR 60.17, WAR 71.62). This setting uses learnable context prompts with coarse class labels and does not use the fine-grained descriptors y' that motivate the method in §3.3.1. Meanwhile, the learnable-descriptor variants (Table 5, settings j–l) obtain UARs of 58.56, 59.23, and 59.75—all below setting h. The only descriptor-based configuration that beats h in UAR is setting e ('Descriptors Add', 60.23), but that configuration has no learnable prompts and is not the one deployed for the SOTA claim. The paper therefore does not demonstrate that fine-grained textual descriptions are the source of improvement; the ablation is internally inconsistent with the abstract and introduction. This is a load-bearing issue for the central claim.
  2. [§4.4.1, Table 3] The text states that 'Based on the UAR, the optimal depth for the fine-grained network is 1, whereas that for the coarse-grained network is 2.' However, the deployed configuration in the SOTA tables is setting e (fine=1, coarse=1), not setting f (fine=1, coarse=2) which has a higher UAR (60.58 vs. 60.17). The selection criterion is unclear: if UAR is the reported metric, setting f should be preferred; if WAR is the reported metric, setting e is preferred (71.62 vs. 71.46). The paper should state which metric was used for model selection and provide per-fold numbers to rule out post-hoc selection on the test folds.
  3. [§4.3, Tables 1–2] No error bars, per-fold breakdowns, or multiple-seed results are reported. The SOTA gaps over DFER-CLIP are 0.56 UAR / 0.37 WAR on DFEW and 0.23 UAR / 0.02 WAR on FERV39k. At this scale, run-to-run variance from random initialization, data augmentation, and hyperparameter selection can easily exceed the reported differences. Given that Tables 3–5 are used to select hyperparameters on the same test folds, the claim of a new state of the art is not statistically substantiated. Reporting per-fold results and at least three seeds for the main configuration would address this.
minor comments (5)
  1. [Abstract] Grammar: 'We incorporate a vision-language pre-trained (VLP) model is integrated' should read 'We incorporate a vision-language pre-trained (VLP) model.'
  2. [§3.3.1, Eq. (6)] The notation P_k = [c]_k^1 [c]_k^2 ... [c]_k^M [Tok(y')]_k is not defined precisely; the superscripts and the token concatenation operation should be clarified, and the equation should specify how the learnable context for the k-th description is initialized.
  3. [§4.3] Typo: 'datset' should be 'dataset.' Also, the sentence 'our method surpasses several state-of-the-art methods' is vague; the specific comparison to DFER-CLIP is given, but the list of methods in Table 2 could be referenced more explicitly.
  4. [§3.2 and §4.4.2] 'transform encoder' appears where 'transformer encoder' is meant. Also, Table 3’s caption omits that the values are on the DFEW 5-fold setting; this should be stated in the caption.
  5. [§3.3.1] The fine-grained descriptors y' are not listed. The paper says they are inherited from [12], but for reproducibility and to assess their discriminative power across the seven emotions, the full set of descriptors should be provided in an appendix or supplementary material.

Circularity Check

0 steps flagged · score 1.0 of 10

No derivation-level circularity; the main caveat is test-fold configuration selection, which is a soundness concern rather than a circular step.

full rationale

The derivation chain is self-contained: frame visual features (Eq. 1), the fine- and coarse-grained temporal encoders (Eqs. 2-5), the text embeddings (Eq. 6), and the visual-prompt alignment (Eqs. 7-10) are all computed from the input video, frozen CLIP encoders, and coarse clip labels; the bag-level prediction in Eq. 11 is trained with cross-entropy against the clip label and evaluated on held-out folds. None of these equations defines the target in terms of itself, and no load-bearing claim reduces to a self-citation: the fine-grained descriptors are inherited from external prior work [12], not from the present authors. The visual-prompt feedback loop is architectural self-attention between visual instances and text features, not a logical circularity. The main caveat is empirical: the 60.17/71.62 headline numbers are the best values among the hyperparameter configurations explored in Tables 3-5 on the same DFEW folds, and Sec. 4.4.1 states the UAR-optimal coarse depth is 2 while the deployed configuration uses coarse depth 1 (Table 3 e vs f). Also, the best prompt configuration in Table 5 ('[Learnable] Class Add') uses coarse class prompts rather than the fine-grained descriptors emphasized in the narrative, weakening the attribution of gains to text guidance. These are benchmark-overfitting and attribution concerns, not a derivation that reduces to its own input, so no circular step is scored.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The framework rests on MIL bag assumptions, CLIP feature adequacy, and hand-selected hyperparameters; the visual-prompt module is the main new architectural entity. No physical entities or new mathematical axioms are introduced.

free parameters (5)
  • Temperature tau = 0.01
    Used in Eq. (8) and (11) for all similarity scaling; set by hand in Sec. 4.2 without sensitivity analysis.
  • Fine-grained window size w and stride s = w=4, s=1
    Selected by ablation on DFEW test folds (Table 4); these directly affect X_fine and final instance features.
  • Temporal depths (fine, coarse) = 1, 1
    Chosen in Table 3 as deployed configuration; note UAR-optimal coarse depth in the same table is 2 (setting f), so the choice is inconsistent with the stated UAR criterion.
  • Learnable context prompts c_i = M learnable vectors before class tokens
    Eq. (6): learnable soft prompts optimized with lr 1e-3; central to text feature X_FP.
  • Training schedule = lr 1e-5/1e-3/1e-2, MultiStepLR milestones 30/40, gamma 0.1
    Section 4.2; standard but hand-set training hyperparameters that affect final accuracy.
assumptions (4)
  • domain assumption Each video bag contains at least one frame expressing the clip-level emotion (MIL positive-bag assumption).
    Section 3.1: all frames in V are treated as expressing y coarsely; if a clip is mislabeled or has no expression, bag-level supervision is corrupted.
  • domain assumption Fine-grained text descriptions y' such as "a smiling mouth, widened eyes" are semantically discriminative for the seven emotions and transfer across DFEW and FERV39k.
    Section 3.3.1: X_FP is built from these descriptions; no per-class descriptor list or validation is provided in this paper.
  • domain assumption Frozen CLIP ViT-B/32 image/text encoders provide adequate frame-level visual features and text alignment for subtle facial dynamics.
    Sections 3.2 and 3.3 use frozen CLIP encoders; the method adds no visual adapter, so failure of CLIP features to capture micro-expressions would cap performance.
  • domain assumption Benchmark labels in DFEW/FERV39k are reliable enough for the reported 5-fold comparisons to be meaningful.
    Section 4.1 describes crowd-annotated labels; the paper does not analyze label noise.
invented entities (1)
  • Visual-prompt-refined label feature (V_P and X_tilde_p)
    purpose: Refines class text embedding with video-specific visual context before similarity-based classification (Eq. 7-10); claimed to improve frame relevance estimation.
    Only evidence is in-paper ablation (Table 5) and qualitative heatmaps; no external falsifiable handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text-guided Weakly Supervised Framework for Dynamic Facial Expression Recognition." pith.science (2026). https://pith.science/paper/BF2I5RX2

@misc{pith2026251110958,
  author       = {Pith},
  title        = {Pith review of: Text-guided Weakly Supervised Framework for Dynamic Facial Expression Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BF2I5RX2}},
  note         = {Machine review of arXiv:2511.10958}
}
read the original abstract

Dynamic facial expression recognition (DFER) aims to identify emotional states by modeling the temporal changes in facial movements across video sequences. A key challenge in DFER is the many-to-one labeling problem, where a video composed of numerous frames is assigned a single emotion label. A common strategy to mitigate this issue is to formulate DFER as a Multiple Instance Learning (MIL) problem. However, MIL-based approaches inherently suffer from the visual diversity of emotional expressions and the complexity of temporal dynamics. To address this challenge, we propose TG-DFER, a text-guided weakly supervised framework that enhances MIL-based DFER by incorporating semantic guidance and coherent temporal modeling. We incorporate a vision-language pre-trained (VLP) model is integrated to provide semantic guidance through fine-grained textual descriptions of emotional context. Furthermore, we introduce visual prompts, which align enriched textual emotion labels with visual instance features, enabling fine-grained reasoning and frame-level relevance estimation. In addition, a multi-grained temporal network is designed to jointly capture short-term facial dynamics and long-range emotional flow, ensuring coherent affective understanding across time. Extensive results demonstrate that TG-DFER achieves improved generalization, interpretability, and temporal sensitivity under weak supervision.

Figures

Figures reproduced from arXiv: 2511.10958 by the authors.

Figure 1
Figure 1. Overview of the proposed text-guided weakly supervised framework for DFER. The weakly super [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of the proposed TG-DFER framework. (a) The overall architecture integrates visual and [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Confusion matrix of our proposed TG-DFER evaluated on 5-fold DFEW (a)-(e) and FERV39k (f). [PITH_FULL_IMAGE:figures/full_fig_p023_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization results of enhanced fine-grained label feature influence in the DFEW dataset. [PITH_FULL_IMAGE:figures/full_fig_p024_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 2 linked inside Pith

  1. [12]

    Prompting visual-language models for dynamic facial expression recognition,

    Z. Zhao, and I. Patras, “Prompting visual-language models for dynamic facial expression recognition,” inBritish Machine Vision Conference, pp. 1–14, 2023

    2023

  2. [1]

    A facial expression emotion recognition based human-robot interaction system,

    Z. Liu, M. Wu, W. Cao, L. Chen, J. Xu, R. Zhang, M. Zhou, and J. Mao, “A facial expression emotion recognition based human-robot interaction system,” IEEE/CAA Journal of Automatica Sinica, vol. 4, no. 4, pp. 668–676, 2017

    2017

  3. [2]

    Schinet: Automatic estimation of symptoms of schizophrenia from facial behaviour analysis,

    M. Bishay, P. Palasek, S. Priebe, and I. Patras, “Schinet: Automatic estimation of symptoms of schizophrenia from facial behaviour analysis,”IEEE Transactions on Affective Computing, vol. 12, no. 4, pp. 949–961, 2019

    2019

  4. [3]

    Towards facial expression analysis in a driver assistance system,

    T. Wilhelm, “Towards facial expression analysis in a driver assistance system,” in Proceedings of the 2019 14th IEEE International Conference on Automatic Face &Gesture Recognition, pp. 1–4, 2019

    2019

  5. [4]

    Suppressing uncertainties for large-scale facial expression recognition,

    K. Wang, X. Peng, J. Yang, S. Lu, and Y . Qiao, “Suppressing uncertainties for large-scale facial expression recognition,” inProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pp. 6897–6906, 2020

    2020

  6. [5]

    Facial expression recognition with visual transformers and attentional selective fusion,

    F. Ma, B. Sun, and S. Li, “Facial expression recognition with visual transformers and attentional selective fusion,”IEEE Transactions on Affective Computing, vol. 14, no. 2, pp. 1236–1248, 2021

    2021

  7. [6]

    Former-dfer: Dynamic facial expression recognition trans- former,

    Z. Zhao, and Q. Liu, “Former-dfer: Dynamic facial expression recognition trans- former,” inProceedings of the 29th ACM International Conference on Multime- dia, pp. 1553–1561, 2021

    2021

  8. [7]

    Dpcnet: Dual path multi-excitation collaborative network for facial expression representation learning in videos,

    Y . Wang, Y . Sun, W. Song, S. Gao, Y . Huang, Z. Chen, W. Ge, and W. Zhang, “Dpcnet: Dual path multi-excitation collaborative network for facial expression representation learning in videos,” inProceedings of the 30th ACM International Conference on Multimedia, pp. 101–110, 2022. 27

    2022

Show all 55 references
  1. [8]

    Expression snippet transformer for robust video-based facial expression recognition,

    Y . Liu, W. Wang, C. Feng, H. Zhang, Z. Chen, and Y . Zhan, “Expression snippet transformer for robust video-based facial expression recognition,”Pattern Recog- nition, vol. 138, pp. 109368, 2023

    2023

  2. [9]

    Spatio-temporal transformer for dynamic facial expres- sion recognition in the wild,

    F. Ma, B. Sun, and S. Li, “Spatio-temporal transformer for dynamic facial expres- sion recognition in the wild,”arXiv preprint arXiv:2205.04749, 2022

    2022 arXiv

  3. [10]

    Logo-former: Local-global spatio-temporal trans- former for dynamic facial expression recognition,

    F. Ma, B. Sun, and S. Li, “Logo-former: Local-global spatio-temporal trans- former for dynamic facial expression recognition,” inICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing, pp. 1–5, 2023

    2023

  4. [11]

    Cliper: A unified vision-language frame- work for in-the-wild facial expression recognition,

    H. Li, H. Niu, Z. Zhu, and F. Zhao, “Cliper: A unified vision-language frame- work for in-the-wild facial expression recognition,” in2024 IEEE International Conference on Multimedia and Expo, 2024

    2024

  5. [13]

    MIDAS: Mixing Ambiguous Data With Soft Labels for Dynamic Facial Expression Recognition,

    R. Kawamura, H. Hayashi, N. Takemura, and H. Nagahara, “MIDAS: Mixing Ambiguous Data With Soft Labels for Dynamic Facial Expression Recognition,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 6552–6562, 2024

    2024

  6. [14]

    Rethinking the learning paradigm for dynamic facial expression recognition,

    H. Wang, B. Li, S. Wu, S. Shen, F. Liu, S. Ding, and A. Zhou, “Rethinking the learning paradigm for dynamic facial expression recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 17958–17968, 2023

    2023

  7. [15]

    Vadclip: Adapting vision-language models for weakly supervised video anomaly detec- tion,

    P. Wu, X. Zhou, G. Pang, L. Zhou, Q. Yan, P. Wang, and Y . Zhang, “Vadclip: Adapting vision-language models for weakly supervised video anomaly detec- tion,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 38, no. 6, pp. 6074–6082, 2024

    2024

  8. [16]

    Dive into ambiguity: La- tent distribution mining and pairwise uncertainty estimation for facial expression 28 recognition,

    J. She, Y . Hu, H. Shi, J. Wang, Q. Shen, and T. Mei, “Dive into ambiguity: La- tent distribution mining and pairwise uncertainty estimation for facial expression 28 recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6248–6257, 2021

    2021

  9. [17]

    Transfer: Learning relation-aware facial expres- sion representations with transformers,

    F. Xue, Q. Wang, and G. Guo, “Transfer: Learning relation-aware facial expres- sion representations with transformers,” inProceedings of the IEEE/CVF Inter- national Conference on Computer Vision, pp. 3601–3610, 2021

    2021

  10. [18]

    Spatial–temporal recurrent neural network for emotion recognition,

    T. Zhang, W. Zheng, Z. Cui, Y . Zong, and Y . Li, “Spatial–temporal recurrent neural network for emotion recognition,”IEEE Transactions on Cybernetics, vol. 49, no. 3, pp. 839–847, 2018

    2018

  11. [19]

    Visual scene-aware hy- brid neural network architecture for video-based facial expression recognition,

    M. K. Lee, D. Y . Choi, D. H. Kim, and B. C. Song, “Visual scene-aware hy- brid neural network architecture for video-based facial expression recognition,” inProceedings of the 2019 14th IEEE International Conference on Automatic Face&Gesture Recognition, pp. 1–8, 2019

    2019

  12. [20]

    Context-aware emotion recog- nition networks,

    J. Lee, S. Kim, S. Kim, J. Park, and K. Sohn, “Context-aware emotion recog- nition networks,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10143–10152, 2019

    2019

  13. [21]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,”Advances in Neural Infor- mation Processing Systems, vol. 30, 2017

    2017

  14. [22]

    Nr-dfernet: Noise-robust network for dynamic facial expression recognition,

    H. Li, M. Sui, and Z. Zhu, “Nr-dfernet: Noise-robust network for dynamic facial expression recognition,”arXiv preprint arXiv:2206.04975, 2022

    2022 arXiv

  15. [23]

    Intensity-aware loss for dynamic facial expression recognition in the wild,

    H. Li, H. Niu, Z. Zhu, and F. Zhao, “Intensity-aware loss for dynamic facial expression recognition in the wild,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, pp. 67–75, 2023

    2023

  16. [24]

    Mae-dfer: Efficient masked autoencoder for self-supervised dynamic facial expression recognition,

    L. Sun, Z. Lian, B. Liu, and J. Tao, “Mae-dfer: Efficient masked autoencoder for self-supervised dynamic facial expression recognition,” inProceedings of the 31st ACM International Conference on Multimedia, pp. 6110–6121, 2023. 29

    2023

  17. [25]

    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, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” inInternational Conference on Machine Learning, pp. 8748–8763, 2021

    2021

  18. [26]

    Scaling up visual and vision-language representation learning with noisy text supervision,

    C. Jia, Y . Yang, Y . Xia, Y . T. Chen, Z. Parekh, H. Pham, Q. Le, Y . Sung, Z. Li, and T. Duerig, “Scaling up visual and vision-language representation learning with noisy text supervision,” inInternational Conference on Machine Learning, pp. 4904–4916, 2021

    2021

  19. [27]

    Zero-shot temporal action detection via vision-language prompting,

    S. Nag, X. Zhu, Y . Z. Song, and T. Xiang, “Zero-shot temporal action detection via vision-language prompting,” inProceedings of the European Conference on Computer Vision, pp. 681–697, 2022

    2022

  20. [28]

    Expanding language-image pretrained models for general video recognition,

    B. Ni, H. Peng, M. Chen, S. Zhang, G. Meng, J. Fu, S. Xiang, and H. Ling, “Expanding language-image pretrained models for general video recognition,” in Proceedings of the European Conference on Computer Vision, pp. 1–18, 2022

    2022

  21. [29]

    Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning,

    H. Luo, L. Ji, M. Zhong, Y . Chen, W. Lei, N. Duan, and T. Li, “Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning,”Neuro- computing, vol. 508, pp. 293–304, 2022

    2022

  22. [30]

    Revisiting temporal modeling for clip-based image-to-video knowledge transferring,

    R. Liu, J. Huang, G. Li, J. Feng, X. Wu, and T. H. Li, “Revisiting temporal modeling for clip-based image-to-video knowledge transferring,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 6555–6564, 2023

    2023

  23. [31]

    Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization,

    Z. Weng X. Yang, A. Li, Z. Wu, and Y . G. Jiang, “Open-vclip: Transforming clip to an open-vocabulary video model via interpolated weight optimization,” in International Conference on Machine Learning, pp. 36978–36989, 2023

    2023

  24. [32]

    X-clip: End-to-end multi- grained contrastive learning for video-text retrieval,

    Y . Ma, G. Xu, X. Sun, M. Yan, J. Zhang, and R. Ji, “X-clip: End-to-end multi- grained contrastive learning for video-text retrieval,” inProceedings of the 30th ACM International Conference on Multimedia, pp. 638–647, 2022. 30

    2022

  25. [33]

    The power of scale for parameter- efficient prompt tuning,

    B. Lester, R. Al-Rfou, and N. Constant, “The power of scale for parameter- efficient prompt tuning,” inProceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pp. 3045–3059, 2021

    2021

  26. [34]

    Visual prompt tuning,

    M. Jia, L. Tang, B. C. Chen, C. Cardie, S. Belongie, B. Hariharan, and S. N. Lim, “Visual prompt tuning,” inProceedings of the European Conference on Computer Vision, pp. 709–727, 2022

    2022

  27. [35]

    Learning to prompt for vision-language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision-language models,”International Journal of Computer Vision, vol. 130, no. 9, pp. 2337– 2348, 2022

    2022

  28. [36]

    Conditional prompt learning for vision- language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Conditional prompt learning for vision- language models,” inProceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pp. 16816–16825, 2022

    2022

  29. [37]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inInternational Conference on Learning ...

    2021

  30. [38]

    Dfew: A large-scale database for recognizing dynamic facial expressions in the wild,

    X. Jiang, Y . Zong, W. Zheng, C. Tang, W. Xia, C. Lu, and J. Liu, “Dfew: A large-scale database for recognizing dynamic facial expressions in the wild,” in Proceedings of the 28th ACM International Conference on Multimedia, pp. 2881– 2889, 2022

    2022

  31. [39]

    Ferv39k: A large-scale multi-scene dataset for facial expression recognition in videos,

    Y . Wang, Y . Sun, Y . Huang, Z. Liu, S. Gao, W. Zhang, W. Ge, and W. Zhang, “Ferv39k: A large-scale multi-scene dataset for facial expression recognition in videos,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 20922–20931, 2022

    2022

  32. [40]

    Learning spatiotem- poral features with 3d convolutional networks,

    D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, “Learning spatiotem- poral features with 3d convolutional networks,” inProceedings of the IEEE Inter- national Conference on Computer Vision, pp. 4489–4497, 2015. 31

    2015

  33. [41]

    Learning spatio-temporal representation with pseudo-3d residual networks,

    Z. Qiu, T. Yao, and T. Mei, “Learning spatio-temporal representation with pseudo-3d residual networks,” inProceedings of the IEEE International Con- ference on Computer Vision, pp. 5533–5541, 2017

    2017

  34. [42]

    Quo vadis, action recognition? a new model and the kinetics dataset,

    J. Carreira, and A. Zisserman, “Quo vadis, action recognition? a new model and the kinetics dataset,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognitionn, pp. 6299–6308, 2017

    2017

  35. [43]

    Can spatiotemporal 3d cnns retrace the his- tory of 2d cnns and imagenet?,

    K. Hara, H. Kataoka, and Y . Satoh, “Can spatiotemporal 3d cnns retrace the his- tory of 2d cnns and imagenet?,” inProceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition, pp. 6546–6555, 2018

    2018

  36. [44]

    A closer look at spatiotemporal convolutions for action recognition,

    D. Tran, H. Wang, L. Torresani, J. Ray, Y . LeCun, and M. Paluri, “A closer look at spatiotemporal convolutions for action recognition,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 6450–6459, 2018

    2018

  37. [45]

    Frame level emotion guided dynamic facial expression recognition with emotion grouping,

    B. Lee, H. Shin, B. Ku, and H. Ko, “Frame level emotion guided dynamic facial expression recognition with emotion grouping,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 5680–5690, 2023

    2023

  38. [46]

    Multiple instance learning for digital pathol- ogy: A review of the state-of-the-art, limitations & future potential,

    M. Gadermayr, and M. Tschuchnig, “Multiple instance learning for digital pathol- ogy: A review of the state-of-the-art, limitations & future potential,”Computer- ized Medical Imaging and Graphics, pp. 102337, 2024

    2024

  39. [47]

    Weakly supervised salient object detection using image labels,

    G. Li, Y . Xie, and L. Lin, “Weakly supervised salient object detection using image labels,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, 2018

    2018

  40. [48]

    Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification,

    H. Zhang, Y . Meng, Y . Zhao, Y . Qiao, X. Yang, S. E. Coupland, and Y . Zheng, “Dtfd-mil: Double-tier feature distillation multiple instance learning for histopathology whole slide image classification,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    2022

  41. [49]

    Multi- ple instance learning for emotion recognition using physiological signals,

    L. Romeo, A. Cavallo, L. Pepa, N. Bianchi-Berthouze, and M. Pontil, “Multi- ple instance learning for emotion recognition using physiological signals,”IEEE Transactions on Affective Computing, vol. 13, no. 1, pp. 389–407, 2019

    2019

  42. [50]

    Learning pain from action unit combina- tions: a weakly supervised approach via multiple instance learning,

    Z. Chen, R. Ansari, and D. J. Wilkie, “Learning pain from action unit combina- tions: a weakly supervised approach via multiple instance learning,”IEEE Trans- actions on Affective Computing, vol. 13, no. 1, pp. 135–146, 2019

    2019

  43. [51]

    Learning multiscale active facial patches for expression analysis,

    L. Zhong, Q. Liu, P. Yang, J. Huang, and D. N. Metaxas, “Learning multiscale active facial patches for expression analysis,”IEEE Transactions on Cybernetics, vol. 145, no. 8, pp. 1499–1510, 2014

    2014

  44. [52]

    Multimodal Prompt Alignment for Facial Ex- pression Recognition,

    F. Ma, Y . He, B. Sun, and S. Li, “Multimodal Prompt Alignment for Facial Ex- pression Recognition,” inProceedings of the IEEE International Conference on Computer Vision, 2025

    2025

  45. [53]

    Facial expression recognition with heatmap neighbor contrastive learning,

    T. Liu, J. Li, J. Wu, B. Du, Y . Zhan, D. Tao, and J. Wan, “Facial expression recognition with heatmap neighbor contrastive learning,”IEEE Transactions on Multimedia, 2025

    2025

  46. [54]

    Confusable facial expression recognition with geometry-aware conditional network,

    T. Liu, J. Li, J. Wu, B. Du, J. Wan, and J. Chang, “Confusable facial expression recognition with geometry-aware conditional network,”Pattern Recognition, vol. 148, pp. 110174, 2024

    2024

  47. [55]

    Cross-domain facial expression recognition via disentangling identity representation,

    T. Liu, J. Li, J. Wu, L. Zhang, S. Zhao, J. Chang, and J. Wan, “Cross-domain facial expression recognition via disentangling identity representation,” inInternational Joint Conferences on Artificial Intelligence, 2023. 33

    2023

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.