Pith. sign in

REVIEW 4 major objections 6 minor 57 references

Learning from Heterogeneity: Generalizing Dynamic Facial Expression Recognition via Distributionally Robust Optimization

T0 review · 4 major / 6 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 A framework called HDF claims that modeling distributional shifts in both time-frequency features and loss optimization improves dynamic facial expression recognition, and reports outperforming prior methods on DFEW and FERV39k.

desk verdict A competent DFER paper with modest single-run gains and a few equation-level loose ends; the SOTA claim is not statistically established, but the assembly is new and the code is released. read the letter →

arxiv 2507.15765 v2 pith:22D4CSUA submitted 2025-07-21 cs.CV cs.AI

Feng-Qi Cui , Anyang Tong , Jinyang Huang , Jie Zhang , Dan Guo , Zhi Liu , Meng Wang This is my paper
classification cs.CVcs.AI
keywords dynamicfacialexpressionrecognitionsampleheterogeneitydistributionallyrobustoptimizationtime-frequencyattentioncontrastivelearninghardsamplesDFEWFERV39k
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 claims that the main obstacle to generalizing dynamic facial expression recognition (DFER) is sample heterogeneity: videos come from different sources, resolutions, and individuals, producing hard samples that destabilize training and blur class boundaries. To counter this, it proposes HDF, a framework with two plug-and-play modules on an X3D video backbone: DAM, a dual-branch time-frequency attention module that adds adversarial and wasserstein-inspired perturbations, and DSM, an optimization module that adaptively reweights classification and contrastive losses based on gradient norm. The paper reports that HDF outperforms state-of-the-art methods on DFEW and FERV39k, with weighted average recall 71.60 and 60.40 on DFEW and 50.30 and 40.49 on FERV39k. If these results hold, the modules would offer a way to improve generalization in heterogeneous video data without dataset-specific tuning.

What carries the argument

The load-bearing mechanism is the pair of modules DAM and DSM. DAM's frequency branch computes DCT features and adds sign-based and wasserstein-guided perturbations to simulate worst-case style shifts, while its temporal branch uses average and max pooling descriptors, a global deviation term approximated as an L2 distance to the sequence mean, and a local temporal difference to build attention maps. DSM's central object is a self-balancing total loss $L_{total} = \lambda_{CE}L_{CE} + \lambda_{SC}L_{IB-D-SC}$ whose weights depend on the gradient norm $g = \|\nabla_w L_{total}\|_2$ through $\lambda_{CE} = \frac{g}{g+1}\alpha$ and $\lambda_{SC} = \frac{1}{g+1}\beta$, with the contrastive term also including gaussian down-weighting of hard negative pairs and an information-bottleneck trace penalty. This gradient-weight coupling is what makes the optimization adaptive; it is also the point where the paper's equations are least explicit.

What would settle it

On DFEW fd5, rerun ablation Setting (c) with the weights in Eq. 16 replaced by constants (for example, $\lambda_{CE} = \lambda_{SC} = 1$) and compare WAR; if the gap between Setting (c) and the baseline disappears, the reported DSM gain comes from the specific weighting scheme. As a second check, instrument the released code to see whether the gradient norm in Eq. 16 is computed with the $\lambda$ weights detached from the computation graph, since the equations as written are self-referential.

Watch

Extended reading notes

Core claim

The central claim is that distributionally robust optimization principles belong inside a DFER model at two levels: feature construction and loss design. DAM's frequency branch treats a DCT-transformed feature as the base distribution and perturbs it with a sign-based adversarial term plus a global-shift term scaled by deviation from the batch mean, approximating the inner maximization of a DRO problem; its temporal branch builds attention from pooled descriptors, a global deviation term (called wasserstein-inspired and computed as an L2 distance to the sequence mean), and a local difference term. DSM then reweights the supervised contrastive loss with a gaussian kernel that down-weights extreme negative pairs, adds an information-bottleneck trace penalty, and scales the classification versus contrastive losses by the current gradient norm so that hard batches emphasize classification. The paper presents ablations showing each branch and each module contributes on DFEW fd5, and per-class numbers show large gains on the least represented emotions, such as disgust rising to 12.41 percent and fear to 41.63 percent.

Load-bearing premise

The adaptive weighting rule computes a gradient length from a loss whose weights depend on that same gradient length; the paper never says the weights are held fixed during the computation, so the rule is only well-defined if that unstated detach step is in the code.

Editorial extensions

If this is right

  • The same training configuration transfers across DFEW and FERV39k, so HDF could be deployed on a new in-the-wild dataset without per-dataset hyperparameter tuning.
  • Per-class gains on rare emotions such as disgust and fear indicate better behavior on underrepresented categories, partially addressing the class imbalance that plagues expression recognition.
  • Because DAM and DSM are described as plug-and-play and independent of the backbone, they can be inserted into existing DFER pipelines beyond the X3D configuration tested here.
  • The flat loss landscape reported for DSM suggests the adaptive scaling acts as a regularizer, which could combine with other sharpness-based training methods for added stability.

Reading between the lines

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

  • Editorial inference: the paper's 'distributionally robust' language is used heuristically, since DAM and DSM approximate DRO with sign perturbations and L2 deviation terms rather than solving a formal worst-case distribution problem.
  • Editorial inference: the gaussian reweighting in the contrastive loss likely behaves like a temperature-adjusted focal loss, so a testable extension is to isolate whether the gain comes from reweighting or from the information-bottleneck trace term.
  • Editorial inference: the gradient-norm scaling depends only on the whole batch, so it cannot distinguish genuinely hard samples from label noise; conditioning scaling on per-sample gradient norms or prediction confidence is a natural next experiment.
  • Editorial inference: the reported numbers are single runs on two benchmarks, so a multi-seed evaluation and ablations on FERV39k would clarify how much of the gain is module-specific versus initialization luck.
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 / 6 minor

Summary. The paper proposes HDF, a framework for dynamic facial expression recognition built on an X3D backbone, with two plug-and-play modules: DAM, a dual-branch time-frequency attention module using DCT, adversarial perturbation, and Wasserstein-inspired temporal regularization; and DSM, a distributionally robust supervised contrastive loss with information-bottleneck regularization and gradient-norm-based adaptive loss scaling. The method is evaluated on DFEW under the 5-fold cross-validation protocol and on FERV39k, reporting WAR/UAR improvements over prior methods (71.60/60.40 on DFEW; 50.30/40.49 on FERV39k). Ablations on DFEW fold 5 attribute gains to each module, and code is released.

Significance. If the reported results are reproducible, HDF offers a practical and modular contribution to DFER: the idea of adapting classification-vs-contrastive loss weights by batch-level gradient sharpness is interesting, and the dual-branch time-frequency attention is a plausible mechanism for handling source-level and individual-level heterogeneity. Strengths of the paper include adherence to standard evaluation protocols, informative component ablations (Tables 3 and 4), a loss-landscape visualization, and a public code release. However, the central claim of 'significantly outperforming' the state of the art is currently supported only by single-run accuracy numbers without variance estimates, and the DSM update rule in Eq. (16) is not fully specified. These issues need to be resolved before the contribution can be fully assessed.

major comments (4)
  1. [§4.2.2, Eq. (16)] The gradient norm g is defined as ||∇_w L_total||_2, but L_total in Eq. (17) itself includes λ_CE and λ_SC, which are the quantities that Eq. (16) is supposed to set. Unless the gradient is computed with the λ weights detached from the computation graph, or a fixed-point/alternating update is specified, Eq. (16) is self-referential and does not define a computable scaling rule. Please state explicitly how g is computed (e.g., stop-gradient on the λ terms) and the order of updates for λ_CE, λ_SC, and the network weights.
  2. [§4.1.3, Eq. (11)] The text states that the fusion weights satisfy the normalization constraint λ_t + λ_s = 1, but Eq. (11) defines each λ as an independent sigmoid output. Sigmoid outputs are not guaranteed to sum to 1, so the constraint is not enforced by the given formulation. Please correct the equation by using a softmax over the two logits or by explicitly normalizing the sigmoid outputs, or state that the implementation performs a subsequent normalization step.
  3. [§5.2, Tables 1 and 2] The abstract and Section 1 claim that HDF 'significantly outperforms' state-of-the-art methods. The reported margins are 0.65–1.51 WAR/UAR, and all numbers are single-run values without standard deviations, confidence intervals, or seed counts for any method. Given run-to-run variance in video transformers and protocol differences (frame sampling, face alignment, augmentation), the word 'significantly' is not supported by the evidence presented. Please add variance estimates from multiple runs, or soften the claim to 'outperforms' with appropriate hedging.
  4. [§5.2 / Abstract] The robustness claim is currently inferred from the same accuracy numbers and from qualitative visualizations (Figs. 3–5); there is no direct distribution-shift or cross-source experiment in the paper. For example, the paper could evaluate on FERV39k scene splits, add a style/noise perturbation test, or train on one source domain and test on another. Without such an experiment, the statement that HDF improves 'robustness to sample heterogeneity' over prior methods is not directly supported.
minor comments (6)
  1. [§4.1.1, Eqs. (6)–(7)] The symbols α and β are used both as learnable perturbation weights in Eq. (6) and as scaling/affine parameters in Eq. (7), while γ is introduced without definition in Eq. (7). Please use distinct symbols for distinct roles.
  2. [§5.1.3] Please state the number of training runs/seeds and the exact software environment (PyTorch version, CUDA version, etc.) to support reproducibility.
  3. [Figure 2 caption] The caption contains a duplicated phrase 'The pipeline of The pipeline of' that should be corrected.
  4. [§4.2.1, Eq. (13)] The robust weight w_ij is defined for negative pairs (i,j), but the denominator in Eq. (13) applies w_{ia} to all candidates a ≠ i. Please clarify whether w is applied to positive pairs as well, and if so, how it is defined for them.
  5. [§5.3.1, Table 3] The text says 'Adding DSM alone leads to even higher gains' but Table 3 shows Setting (c) (DSM alone) at 71.79 WAR, which is higher than Setting (b) (DAM alone) at 69.73, but lower than Setting (d). Please clarify the comparison explicitly, since the sentence could be read as claiming DSM alone is the largest contributor overall.
  6. [§5.1.3] The citation to prior work [43] for setting τ=0.07 and η=0.2 seems to refer to a NeurIPS paper on contrastive learning and distributionally robust optimization; please verify that the specific hyperparameter values come from that reference, or provide a more direct citation for the chosen values.

Circularity Check

1 steps flagged · score 4.0 of 10

DSM's adaptive loss weights are defined through the gradient of the very loss they define; the benchmark SOTA claim is external, so only partial circularity.

  1. self definitional [Section 4.2.2, Eq. (16) and Section 4.2.3, Eq. (17)]
    "Let the gradient norm of the total loss be: g = ||∇_w L_total||_2, then compute the dynamic weights for each loss term as: λ_CE = g/(g+1)·α, λ_SC = 1/(g+1)·β, where α and β are base scaling coefficients. ... The overall objective of DSM is defined as follows: L_total = λ_CE·L_CE + λ_SC·L_IB-D-SC."

    λ_CE and λ_SC are functions of g, while g is the norm of the gradient of L_total, and L_total is itself defined as the weighted sum using those same λ values. Thus Eqs. (16) and (17) form a self-referential system: the weights determine the loss, the loss determines the gradient, and the gradient determines the weights. The paper neither detaches λ when computing the gradient nor specifies a fixed-point iteration or use of a previous step's g. Without such an unstated intervention, the 'adaptive' scaling is not a well-defined function of the optimization state. Since DSM is a central contribution and its ablation gain is the largest in Table 3, the claimed optimization mechanism is not supported by the equations as written.

full rationale

The central SOTA claim is externally grounded and not circular: Tables 1 and 2 compare WAR/UAR against previously published numbers on DFEW and FERV39k, with no parameter fitted to the test metric. The related-work self-citations are contextual and not load-bearing; there is no invoked uniqueness theorem and no fitted input renamed as a prediction. The one genuine circular step is internal to DSM: the loss-scaling weights are defined through the gradient norm of a total loss that is itself defined with those weights. This makes the DSM component self-referential as written, though an implementation could resolve it by detaching the weights when computing the gradient. Because the benchmark result does not reduce by construction, the circularity is partial rather than total, giving a score of 4 rather than 6 or higher.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on a small number of hyperparameters (tau, eta, DSM base coefficients), several ad hoc approximations that borrow the language of DRO, Wasserstein distance, and information bottleneck, and the unstated comparability of single-run benchmark numbers. No new physical entities are introduced.

free parameters (4)
  • temperature tau = 0.07
    Used in supervised contrastive loss (Eq 5, 13); set based on prior work [43], not derived or tuned in this paper.
  • Gaussian reweighting center eta = 0.2
    Controls down-weighting of negative pairs in Eq 13; set based on prior work [43].
  • DSM base scaling coefficients alpha and beta = not reported
    Eq 16 scales classification and contrastive losses; the paper gives no values or search range.
  • DAM learnable weights alpha, beta, gamma, delta = learned
    Eqs 6, 7, 9 learn perturbation strength and gating weights; they are trained parameters rather than chosen constants, but their initialization/range is not specified.
assumptions (6)
  • domain assumption DCT decomposition of video features captures source-level style and quality variation relevant to DFER
    Sec 4.1.1 builds the frequency branch on this premise without independent validation.
  • ad hoc to paper The L2 distance between a frame descriptor and the sequence mean approximates the Wasserstein-regularized temporal deviation
    Sec 4.1.2 defines W_global as an L2 norm while calling it Wasserstein-inspired; no optimal transport derivation is given.
  • ad hoc to paper Gaussian similarity reweighting implements a distributionally robust contrastive objective
    Sec 4.2.1 asserts DRO-enhanced SCL; the formula is heuristic reweighting without an uncertainty set or minimax formulation.
  • ad hoc to paper Trace of feature covariance approximates the information bottleneck penalty
    Sec 4.2.1 sets L_IB = Tr(Cov(Z)); standard information bottleneck uses mutual information, so this is a loose proxy.
  • domain assumption Gradient norm of the total loss is a reliable batch-level difficulty signal for loss weighting
    Sec 4.2.2 uses g = ||grad_w L_total||_2 to set lambda_CE and lambda_SC; supported only by analogy to SAM, and the equation is self-referential unless weights are detached.
  • domain assumption Single-run comparisons against previously published numbers are sufficient to establish superiority
    Section 5.2 compares WAR/UAR without error bars or significance tests.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning from Heterogeneity: Generalizing Dynamic Facial Expression Recognition via Distributionally Robust Optimization." pith.science (2026). https://pith.science/paper/22D4CSUA

@misc{pith2026250715765,
  author       = {Pith},
  title        = {Pith review of: Learning from Heterogeneity: Generalizing Dynamic Facial Expression Recognition via Distributionally Robust Optimization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/22D4CSUA}},
  note         = {Machine review of arXiv:2507.15765}
}
read the original abstract

Dynamic Facial Expression Recognition (DFER) plays a critical role in affective computing and human-computer interaction. Although existing methods achieve comparable performance, they inevitably suffer from performance degradation under sample heterogeneity caused by multi-source data and individual expression variability. To address these challenges, we propose a novel framework, called Heterogeneity-aware Distributional Framework (HDF), and design two plug-and-play modules to enhance time-frequency modeling and mitigate optimization imbalance caused by hard samples. Specifically, the Time-Frequency Distributional Attention Module (DAM) captures both temporal consistency and frequency robustness through a dual-branch attention design, improving tolerance to sequence inconsistency and visual style shifts. Then, based on gradient sensitivity and information bottleneck principles, an adaptive optimization module Distribution-aware Scaling Module (DSM) is introduced to dynamically balance classification and contrastive losses, enabling more stable and discriminative representation learning. Extensive experiments on two widely used datasets, DFEW and FERV39k, demonstrate that HDF significantly improves both recognition accuracy and robustness. Our method achieves superior weighted average recall (WAR) and unweighted average recall (UAR) while maintaining strong generalization across diverse and imbalanced scenarios. Codes are released at https://github.com/QIcita/HDF_DFER.

Figures

Figures reproduced from arXiv: 2507.15765 by the authors.

Figure 1
Figure 1. Due to a series of Sample Heterogeneity problems, [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. An overview of the proposed HDF. (a) The overall processing flow of HDF. (b) The pipeline of The pipeline of Time￾Frequency Distributional Attention (DAM) in HDF. (c) The sketch of Distribution-aware Scaling (DSM). attended sequence representation is then fused with the original temporal features as: 𝑥𝑡 = 𝑥 · 𝜎(TA) + 𝑥, (10) where 𝜎(·) denotes the sigmoid function. In summary, the temporal branch improves sequence-l… view at source ↗
Figure 3
Figure 3. Visualization of the learned feature maps. There are three sequences are presented, which including the issues of the [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustration of feature distribution learned by the baseline and our proposed method in [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The visualization of separate loss landscape on [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 53 canonical work pages

  1. [1]

    Joao Carreira and Andrew Zisserman. 2017. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition

    work page 2017

  2. [2]

    Dongliang Chen, Guihua Wen, Pei Yang, Huihui Li, Chuyun Chen, and Bao Wang

    work page

  3. [3]

    Haodong Chen, Haojian Huang, Junhao Dong, Mingzhe Zheng, and Dian Shao

  4. [4]

    Junkai Chen, Zenghai Chen, Zheru Chi, and Hong Fu. 2018. Facial Expression Recognition in Video with Multiple Feature Fusion.IEEE Transactions on Affective Computing (2018)

    work page 2018

  5. [5]

    In Proceedings of the 32nd ACM International Conference on Multimedia (MM ’24)

    FineCLIPER: Multi-modal Fine-grained CLIP for Dynamic Facial Expres- sion Recognition with AdaptERs. In Proceedings of the 32nd ACM International Conference on Multimedia (MM ’24)

    work page

  6. [6]

    Wheidima Carneiro de Melo, Eric Granger, and Abdenour Hadid. 2022. A Deep Multiscale Spatiotemporal Network for Assessing Depression From Facial Dy- namics. IEEE Transactions on Affective Computing (2022)

    work page 2022

  7. [7]

    Xi Chen, Simai He, Bo Jiang, Christopher Thomas Ryan, and Teng Zhang. 2021. The Discrete Moment Problem with Nonconvex Shape Constraints. (2021)

    work page 2021

  8. [8]

    Christoph Feichtenhofer. 2020. X3D: Expanding Architectures for Efficient Video Recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2020

Show all 57 references
  1. [9]

    John Duchi and Hongseok Namkoong. 2019. Variance-based regularization with convex objectives. Journal of Machine Learning Research (2019)

    2019

  2. [10]

    Yu Gu, Xiang Zhang, Huan Yan, Jingyang Huang, Zhi Liu, Mianxiong Dong, and Fuji Ren. 2023. WiFE: WiFi and Vision Based Unobtrusive Emotion Recognition via Gesture and Facial Expression. IEEE Transactions on Affective Computing 14, 4 (2023), 2567–2581

    2023

  3. [11]

    Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. 2021. Sharpness-aware Minimization for Efficiently Improving Generalization. In In- ternational Conference on Learning Representations

    2021

  4. [12]

    Dan Guo, Wengang Zhou, Houqiang Li, and Meng Wang. 2018. Hierarchical LSTM for sign language translation. In Proceedings of the AAAI conference on artificial intelligence, Vol. 32

    2018

  5. [13]

    Dan Guo, Kun Li, Bin Hu, Yan Zhang, and Meng Wang. 2024. Benchmarking micro-action recognition: Dataset, methods, and applications. IEEE Transactions on Circuits and Systems for Video Technology 34, 7 (2024), 6238–6252

    2024

  6. [14]

    Jingjing Hu, Dan Guo, Kun Li, Zhan Si, Xun Yang, Xiaojun Chang, and Meng Wang. 2025. Unified static and dynamic network: efficient temporal filtering for video grounding. IEEE Transactions on Pattern Analysis and Machine Intelligence (2025)

    2025

  7. [15]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  8. [16]

    Jinyang Huang, Bin Liu, Chenglin Miao, Xiang Zhang, Jiancun Liu, Lu Su, Zhi Liu, and Yu Gu. 2023. Phyfinatt: An undetectable attack framework against phy layer fingerprint-based wifi authentication. IEEE Transactions on Mobile Computing (2023)

    2023

  9. [17]

    Jinyang Huang, Jia-Xuan Bai, Xiang Zhang, Zhi Liu, Yuanhao Feng, Jianchun Liu, Xiao Sun, Mianxiong Dong, and Meng Li. 2024. Keystrokesniffer: An off-the-shelf smartphone can eavesdrop on your privacy from anywhere. IEEE Transactions on Information Forensics and Security (2024)

    2024

  10. [18]

    Xingxun Jiang, Yuan Zong, Wenming Zheng, Chuangao Tang, Wanchuang Xia, Cheng Lu, and Jiateng Liu. 2020. DFEW: A Large-Scale Database for Recog- nizing Dynamic Facial Expressions in the Wild. In Proceedings of the 28th ACM International Conference on Multimedia

    2020

  11. [19]

    Ruomin Huang, Jiawei Huang, Wenjie Liu, and Hu Ding. 2022. Coresets for wasserstein distributionally robust optimization problems. Advances in Neural Information Processing Systems (2022)

    2022

  12. [20]

    Jiyoung Lee, Sunok Kim, Seungryong Kim, and Kwanghoon Sohn. 2020. Multi- Modal Recurrent Attention Networks for Facial Expression Recognition. IEEE Transactions on Image Processing (2020)

    2020

  13. [21]

    Binbin Jin, Defu Lian, Zheng Liu, Qi Liu, Jianhui Ma, Xing Xie, and Enhong Chen

  14. [22]

    Hanting Li, Hongjing Niu, Zhaoqing Zhu, and Feng Zhao. 2024. CLIPER: A Unified Vision-Language Framework for In-the-Wild Facial Expression Recognition. In 2024 IEEE International Conference on Multimedia and Expo (ICME)

    2024

  15. [23]

    Mengke Li, Ye Liu, Yang Lu, Yiqun Zhang, Yiu ming Cheung, and Hui Huang

  16. [24]

    Hanting Li, Hongjing Niu, Zhaoqing Zhu, and Feng Zhao. 2023. Intensity-aware loss for dynamic facial expression recognition in the wild. AAAI

    2023

  17. [25]

    Feng Liu, Hanyang Wang, and Siyuan Shen. 2025. Robust Dynamic Facial Expres- sion Recognition. IEEE Transactions on Biometrics, Behavior, and Identity Science (2025)

    2025

  18. [26]

    Hanwei Liu, Huiling Cai, Qingcheng Lin, Xuefeng Li, and Hui Xiao. 2023. Learn- ing from More: Combating Uncertainty Cross-multidomain for Facial Expression Recognition. In Proceedings of the 31st ACM International Conference on Multime- dia (MM ’23). Association for Computing ...

    2023

  19. [27]

    In The Thirty-eighth Annual Conference on Neural Information Processing Systems

    Improving Visual Prompt Tuning by Gaussian Neighborhood Minimization for Long-Tailed Visual Recognition. In The Thirty-eighth Annual Conference on Neural Information Processing Systems

  20. [28]

    Shan Li and Weihong Deng. 2020. Deep facial expression recognition: A survey. IEEE transactions on affective computing (2020)

    2020

  21. [29]

    Fuyan Ma, Bin Sun, and Shutao Li. 2023. Facial Expression Recognition With Visual Transformers and Attentional Selective Fusion. IEEE Transactions on Affective Computing (2023)

    2023

  22. [30]

    Fuyan Ma, Bin Sun, and Shutao Li. 2023. Logo-Former: Local-Global Spatio- Temporal Transformer for Dynamic Facial Expression Recognition. IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP) (2023)

    2023

  23. [31]

    Yuanyuan Liu, Yuxuan Huang, Shuyang Liu, Yibing Zhan, Zijing Chen, and Zhe Chen. 2024. Open-Set Video-based Facial Expression Recognition with Human Expression-sensitive Prompting. In Proceedings of the 32nd ACM International Conference on Multimedia (MM ’24)

    2024

  24. [32]

    Fuyan Ma, Bin Sun, and Shutao Li. 2022. Spatio-Temporal Transformer for Dynamic Facial Expression Recognition in the Wild. (2022)

    2022

  25. [33]

    Aman Sinha, Hongseok Namkoong, and John Duchi. 2018. Certifiable Distribu- tional Robustness with Principled Adversarial Training. In International Confer- ence on Learning Representations

    2018

  26. [34]

    Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri

  27. [35]

    Paul Michel, Tatsunori Hashimoto, and Graham Neubig. 2022. Distributionally Robust Models with Parametric Likelihood Ratios. In International Conference on Learning Representations

    2022

  28. [36]

    Zequn Qin, Pengyi Zhang, Fei Wu, and Xi Li. 2021. FcaNet: Frequency Channel Attention Networks. In IEEE/CVF International Conference on Computer Vision (ICCV)

    2021

  29. [37]

    Hanyang Wang, Bo Li, Shuang Wu, Siyuan Shen, Feng Liu, Shouhong Ding, and Aimin Zhou. 2023. Rethinking the Learning Paradigm for Dynamic Facial Expression Recognition. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2023

  30. [38]

    Ruiqi Wang, Jinyang Huang, Jie Zhang, Xin Liu, Xiang Zhang, Zhi Liu, Peng Zhao, Sigui Chen, and Xiao Sun. 2024. FacialPulse: An Efficient RNN-based Depression Detection via Temporal Facial Landmarks. In ACM Multimedia 2024

    2024

  31. [39]

    Rui Wang and Xiao Sun. 2023. Dynamic Facial Expression Recognition Based on Vision Transformer with Deformable Module. In 2023 IEEE International Conference on Systems, Man, and Cybernetics (SMC)

    2023

  32. [40]

    Md Azher Uddin, Joolekha Bibi Joolee, and Young-Koo Lee. 2022. Depression Level Prediction Using Deep Spatiotemporal Features and Multilayer Bi-LTSM. IEEE Transactions on Affective Computing (2022)

    2022

  33. [41]

    Laurens van der Maaten and Geoffrey Hinton. 2008. Visualizing Data using t-SNE. Journal of Machine Learning Research (2008)

    2008

  34. [42]

    Chenwang Wu, Defu Lian, Yong Ge, Zhihao Zhu, and Enhong Chen. 2023. Influence-Driven Data Poisoning for Robust Recommender Systems. IEEE Trans- actions on Pattern Analysis and Machine Intelligence (2023)

    2023

  35. [43]

    Junkang Wu, Jiawei Chen, Jiancan Wu, Wentao Shi, Xiang Wang, and Xiangnan He. 2023. Understanding Contrastive Learning via Distributionally Robust Opti- mization. In Thirty-seventh Conference on Neural Information Processing Systems . MM ’25, October 27–31, 2025, Dublin, Irela...

    2023

  36. [44]

    Yi Wu, Shangfei Wang, and Yanan Chang. 2023. Patch-Aware Representation Learning for Facial Expression Recognition. In Proceedings of the 31st ACM Inter- national Conference on Multimedia (MM ’23)

    2023

  37. [45]

    Yan Wang, Yixuan Sun, Yiwen Huang, Zhongying Liu, Shuyong Gao, Wei Zhang, Weifeng Ge, and Wenqiang Zhang. 2022. FERV39k: A Large-Scale Multi-Scene Dataset for Facial Expression Recognition in Videos. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2022

  38. [46]

    Yanan Wang, Jianming Wu, and Keiichiro Hoashi. 2019. Multi-attention fusion network for video-based emotion recognition. In International Conference on Multimodal Interaction

    2019

  39. [47]

    Ziyang Zhang, Xiang Tian, Yuan Zhang, Kailing Guo, and Xiangmin Xu. 2024. Label-Guided Dynamic Spatial-Temporal Fusion for Video-Based Facial Expres- sion Recognition. IEEE Transactions on Multimedia (2024)

    2024

  40. [48]

    Guoying Zhao and Matti Pietikainen. 2007. Dynamic Texture Recognition Using Local Binary Patterns with an Application to Facial Expressions.IEEE Transactions on Pattern Analysis and Machine Intelligence (2007)

    2007

  41. [49]

    Zengqun Zhao and Qingshan Liu. 2021. Former-DFER: Dynamic Facial Expression Recognition Transformer. InProceedings of the 29th ACM International Conference on Multimedia (MM ’21)

    2021

  42. [50]

    Tong Zhang, Wenming Zheng, Zhen Cui, Yuan Zong, and Yang Li. 2019. Spa- tial–Temporal Recurrent Neural Network for Emotion Recognition. IEEE Trans- actions on Cybernetics (2019)

    2019

  43. [51]

    Xiaoqin Zhang, Min Li, Sheng Lin, Hang Xu, and Guobao Xiao. 2024. Transformer- Based Multimodal Emotional Perception for Dynamic Facial Expression Recogni- tion in the Wild. IEEE Transactions on Circuits and Systems for Video Technology (2024)

    2024

  44. [55]

    Jinxing Zhou, Dan Guo, and Meng Wang. 2022. Contrastive positive sample propagation along the audio-visual event line. IEEE Transactions on Pattern Analysis and Machine Intelligence 45, 6 (2022), 7239–7257

    2022

  45. [56]

    Hao Zhu, Wenping Ma, Lingling Li, Licheng Jiao, Shuyuan Yang, and Biao Hou

  46. [57]

    Information Fusion 58 (2020), 116–131

    A Dual–Branch Attention fusion deep network for multiresolution re- mote–Sensing image classification. Information Fusion 58 (2020), 116–131

    2020

  47. [2015]

    In Proceedings of the IEEE international conference on computer vision

    Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision . 4489–4497

  48. [2020]

    Advances in Neural Information Processing Systems (2020)

    Sampling-decomposable generative adversarial recommender. Advances in Neural Information Processing Systems (2020)

    2020

  49. [2024]

    IEEE Transactions on Circuits and Systems for Video Technology (2024)

    CFAN-SDA: Coarse-Fine Aware Network With Static-Dynamic Adaptation for Facial Expression Recognition in Videos. IEEE Transactions on Circuits and Systems for Video Technology (2024)

    2024

Pith tools

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