Pith. sign in

REVIEW 4 major objections 3 minor 76 references

Rethinking Occlusion in FER: A Semantic-Aware Perspective and Go Beyond

T0 review · 4 major / 3 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 The paper claims that facial expression recognition under occlusion improves when dense segmentation maps and sparse landmarks are fused as semantic priors, reaching state-of-the-art accuracy on three benchmarks.

desk verdict A well-ablated FER architecture with a plausible but under-evidenced SOTA claim; the new dataset and variance numbers need to ship before the headline results are credible. read the letter →

arxiv 2507.15401 v3 pith:YKZOBLJL submitted 2025-07-21 cs.CV

Huiyu Zhai , Xingxing Yang , Yalan Ye , Chenyang Li , Bin Fan , Changze Li This is my paper · ORCID
classification cs.CV
keywords FacialexpressionrecognitionOcclusionrobustnessSemanticsegmentationpriorlandmarksMulti-scalecross-attentionfusionOcclu-FERdatasetClassimbalancelossSpatially-adaptivenormalization
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 existing facial expression recognition models fail under occlusion because they treat the face as raw pixels and lack high-level semantic knowledge about which facial parts should be present. ORSANet supplies that knowledge through two external priors: dense pixel-level segmentation maps that say what each facial region is, and sparse facial landmarks that anchor the geometry of the face. A Multi-scale Cross-interaction Module fuses the two priors with image features, and a Dynamic Adversarial Repulsion Enhancement Loss (DARELoss) pushes easily confused expression classes apart. On RAF-DB, AffectNet, and the new Occlu-FER benchmark, the method reports top accuracy, with the smallest drop under artificial occlusion. If the claim is right, occlusion resilience in FER comes from explicit semantic structure rather than from larger or deeper backbones.

What carries the argument

The load-bearing mechanism is the pairing of a dense semantics prior with a sparse geometric prior. The dense prior is a pixel-level facial segmentation map produced by a frozen pre-trained parsing network; it enters through the Spatial-Semantic Guidance Module, where a dual-stage SPADE-style normalization uses the map to compute per-pixel scale and shift for the image features. The sparse prior is a set of facial landmarks from a frozen detector; in the Multi-scale Cross-interaction Module the landmarks become query vectors in a cross-attention over the semantics-enhanced image features, and are then reintegrated with an adaptive scaling factor. DARELoss closes the loop by taking the logits of the most competitive negative class, scaling them by the target class's predicted confidence, and using them in a binary repulsion term, widening the decision boundary around ambiguous expressions.

What would settle it

Take the Occlu-FER validation set and replace each predicted segmentation map with a deliberately corrupted version, such as injecting label noise inside the occluded region or swapping in a map from a different face. If ORSANet's accuracy stays close to 68.07%, the dense semantic prior is not the source of the gain; if accuracy drops sharply, the prior is doing the claimed work.

Watch

Extended reading notes

Core claim

The central claim is that explicitly injecting semantic structure makes occlusion-robust facial expression recognition possible with a standard backbone. ORSANet takes segmentation maps from a frozen pre-trained face-parsing network and uses them, through a spatially-adaptive normalization module, to modulate the image features so the network attends to semantically plausible facial regions even when parts are covered. Landmarks from a frozen detector are then crossed with the enhanced features in a multi-scale attention module, which the paper says separates expression-related cues from identity, gender, and illumination noise. A confidence-aware loss suppresses the most confusing non-target class with a dynamic margin to handle ambiguous expressions. The paper reports 92.28% on RAF-DB, 66.69% on AffectNet 7-class, 62.95% on AffectNet 8-class, and 68.07% on the new Occlu-FER set, and shows that accuracy degrades less than comparison methods as artificial occlusion increases from 10% to 30%.

Load-bearing premise

The method depends on the frozen segmentation network returning correct semantic maps for occluded and low-quality faces; when those maps are wrong, the enhanced features are wrong and the claimed advantage collapses.

Editorial extensions

If this is right

  • Occlusion robustness in FER can be improved without redesigning the backbone; frozen semantic priors added at training time are sufficient to raise accuracy on occluded faces.
  • Artificial random erasing is a weaker training proxy than real semantic occlusion, and the new Occlu-FER benchmark gives future methods a realistic testbed with eight basic expression classes.
  • Because the face-parsing and landmark detectors are frozen, ORSANet's ceiling is set by their quality; better segmentation models should translate directly into better FER accuracy under occlusion.
  • DARELoss is architecture-agnostic, so the paper's reported gains on hyperspectral image classification and CIFAR-10 imply the loss transfers beyond FER to general classification tasks.

Reading between the lines

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

  • A natural extension the paper leaves implicit is to fine-tune or jointly train the face parser on occluded faces; the frozen-prior design suggests that adapting the parser to occlusion could compound ORSANet's gains.
  • Because Occlu-FER contains contextual cues such as body gestures and scene elements, part of the benchmark score may reflect context recognition rather than pure facial occlusion handling; a face-crop-only evaluation would separate those factors.
  • DARELoss can be read as confidence-weighted hard-negative mining; applying it to long-tailed classification with per-class confidence estimates is a direct test the paper does not run.
  • The paper's own limitation section concedes that low-quality or anomalous faces can destabilize feature extraction when the segmentation network errs, so the method's robustness on exactly the inputs it targets is conditional on segmentation quality.
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 / 3 minor

Summary. The manuscript presents ORSANet, a facial expression recognition (FER) method designed to improve occlusion robustness by combining two auxiliary priors: dense semantic segmentation maps from a frozen SegFace model and sparse facial landmarks from a frozen MobileFaceNet detector. The dense prior is injected through a Spatial-Semantic Guidance Module built on SPADE layers, and the landmarks are fused with the resulting semantics-enhanced features through a Multi-scale Cross-interaction Module. The paper also introduces a Dynamic Adversarial Repulsion Enhancement Loss (DARELoss) intended to enlarge decision margins for confusable classes, and constructs a new occlusion-oriented benchmark called Occlu-FER. Experiments are reported on RAF-DB, AffectNet 7/8-class, artificial occlusion versions of RAF-DB, and the new Occlu-FER dataset, with claims of state-of-the-art accuracy (92.28% on RAF-DB, 66.69% on AffectNet 7-class, 62.95% on AffectNet 8-class, 68.07% on Occlu-FER).

Significance. If fully substantiated, the paper would make a useful contribution to occlusion-robust FER: the architecture is clearly specified, the component-wise ablations in Table 6 isolate the effect of each module, the code is publicly available, and the proposed Occlu-FER dataset could fill a real gap in evaluating real-world occlusions. The loss generalization experiments in Tables 7-8, while peripheral, suggest DARELoss may have broader applicability. However, the central empirical claim is currently under-supported: the reported advantages over prior state-of-the-art methods are very small, all results appear to be single runs with no variance or significance testing, and the new benchmark is unreleased and constructed by the same group. The acknowledged limitation in Section 5 additionally raises a correctness-risk concern about the very inputs the method targets. These issues prevent me from accepting the SOTA claim as established.

major comments (4)
  1. [§4.3, Tables 2-4] All accuracy numbers appear to be from a single training run; the paper reports no standard deviations, no number of seeds, and no significance tests. On RAF-DB the reported gain over POSTER V2 is 0.07 percentage points (92.28 vs. 92.21), and on AffectNet 7-class it is 0.46 points (66.69 vs. 66.23); these margins are within the seed-to-seed variation typical of FER training, and the comparison uses previously published numbers whose preprocessing and protocol may not be identical. This does not establish the central "state-of-the-art" claim; repeated runs with a matched protocol and significance testing, or at least explicit variance reporting, are needed.
  2. [§4.1 and §4.3 (Occlu-FER)] The new Occlu-FER benchmark is a main contribution and is used in Tables 4 and 6 to demonstrate the method's occlusion robustness, but the dataset is not released and the description in §4.1 gives no annotation protocol, no annotator agreement, no inclusion/exclusion criteria, and no detail on the internet-collected images (source, consent, license). Because the dataset is constructed by the same group and is not externally verifiable, the reported gains on it (e.g., 68.07 vs. 67.61 for CLIPER in Table 4) cannot be checked. The dataset and its annotation materials should be released, or the Occlu-FER-based claims should be substantially weakened.
  3. [§5, Limitation] The Limitation section states that low-quality or anomalous facial samples can cause instability in feature extraction because of segmentation errors. This directly affects the central mechanism, since ORSANet's occlusion robustness depends on dense segmentation priors, and the occluded/low-quality regime is exactly the paper's target. The manuscript does not quantify how often segmentation maps are wrong on Occlu-FER or RAF-DB (occlu), nor does it compare against a baseline without the prior in those failure cases. A quantitative robustness analysis is needed, for example accuracy split by segmentation quality or sensitivity to corrupted segmentation maps, to establish that the benefit is not limited to inputs where the frozen segmenter already succeeds.
  4. [§4.3, Table 5] The complexity comparison is unfair: the note says only ORSANet's trainable part is counted, but the frozen SegFace and MobileFaceNet models are executed at inference to produce the priors and therefore consume parameters, FLOPs, and memory. Reporting 60.2M parameters and 6.9G FLOPs while other rows report full-model costs understates ORSANet's true inference cost. The table should report full inference cost including both frozen auxiliary networks, or state clearly that the comparison is for trainable parameters only and not a full-system comparison.
minor comments (3)
  1. [§3.1, Eq. (2)] Equation (2) integrates over dX_N2G, but the variable X_N2G is never defined; the integrand uses X_N2S and X_N2L. Please correct the integration variable or define X_N2G.
  2. [§4.3, Table 4] The text reports 68.05% for ORSANet on Occlu-FER and says it outperforms all compared methods "except CLIPER", but Table 4 reports 68.07% and CLIPER at 67.61%, so ORSANet is ahead of CLIPER. The sentence and the earlier "RAD-DB" typo should be corrected.
  3. [§6, Conclusion] DARELoss is misspelled as "DRAELoss" in the conclusion; also, the claims about DARELoss generality in Tables 7-8 would benefit from multi-seed results and significance tests, as the gains are small (e.g., 0.85 percentage points on CrossViT with CIFAR-10).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central SOTA claims rest on external public benchmarks, and the self-created Occlu-FER benchmark is a reproducibility limitation rather than a circular reduction.

full rationale

The paper's main claims are empirical SOTA accuracies on RAF-DB and AffectNet, which are compared against previously published results from independent groups; no equation or fitted parameter forces those numbers to equal the paper's own inputs. The dense-semantics prior comes from a frozen, externally trained SegFace model and the sparse prior from a frozen MobileFaceNet detector (Section 4.2), and the fusion is specified by the trainable equations of SSGM and MCM (Eqs. 4-9), whose output is not defined as the segmentation map or landmark map themselves. DARELoss (Eqs. 10-12) is a training objective computed from logits and is ablated, not a fitted value renamed as a prediction. The author self-citations ([61], [62]) are used only as architecture and latent-prior references and are not load-bearing: the MCM design is fully described and ablated in Section 4.4. The self-constructed Occlu-FER dataset is unreleased, which is an external-validation and reproducibility concern, and Section 5 explicitly acknowledges that segmentation errors on low-quality or anomalous faces can destabilize feature extraction; however, these are verification weaknesses, not cases where a reported result holds by definition. Therefore no significant circularity is present.

Assumptions & free parameters 2 free parameters · 2 assumptions · 0 invented entities

The central claim relies on two domain assumptions: the reliability of the frozen SegFace segmentation prior, and the quality and representativeness of the self-constructed Occlu-FER dataset. The only hand-chosen free parameters are the loss weights, set to 1 and 0.1. No new theoretical entities, particles, or forces are introduced.

free parameters (2)
  • lambda_1 = 1
    Trade-off weight for cross-entropy loss in Eq. 12, set by hand.
  • lambda_2 = 0.1
    Trade-off weight for DARELoss in Eq. 12, set by hand.
assumptions (2)
  • domain assumption Pre-trained SegFace segmentation model produces reliable dense semantic maps for occluded faces.
    Section 4.2 freezes SegFace weights; SSGM uses these maps as the latent variable Z. If segmentation fails under occlusion, the enhancement is corrupted, a limitation acknowledged in Section 5.
  • domain assumption Occlu-FER is a representative, correctly labeled sample of real-world occlusion.
    Section 4.1 describes construction from public datasets and internet images but gives no annotation protocol, source IDs, or release, so label quality and representativeness are unverified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Occlusion in FER: A Semantic-Aware Perspective and Go Beyond." pith.science (2026). https://pith.science/paper/YKZOBLJL

@misc{pith2026250715401,
  author       = {Pith},
  title        = {Pith review of: Rethinking Occlusion in FER: A Semantic-Aware Perspective and Go Beyond},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YKZOBLJL}},
  note         = {Machine review of arXiv:2507.15401}
}
read the original abstract

Facial expression recognition (FER) is a challenging task due to pervasive occlusion and dataset biases. Especially when facial information is partially occluded, existing FER models struggle to extract effective facial features, leading to inaccurate classifications. In response, we present ORSANet, which introduces the following three key contributions: First, we introduce auxiliary multi-modal semantic guidance to disambiguate facial occlusion and learn high-level semantic knowledge, which is two-fold: 1) we introduce semantic segmentation maps as dense semantics prior to generate semantics-enhanced facial representations; 2) we introduce facial landmarks as sparse geometric prior to mitigate intrinsic noises in FER, such as identity and gender biases. Second, to facilitate the effective incorporation of these two multi-modal priors, we customize a Multi-scale Cross-interaction Module (MCM) to adaptively fuse the landmark feature and semantics-enhanced representations within different scales. Third, we design a Dynamic Adversarial Repulsion Enhancement Loss (DARELoss) that dynamically adjusts the margins of ambiguous classes, further enhancing the model's ability to distinguish similar expressions. We further construct the first occlusion-oriented FER dataset to facilitate specialized robustness analysis on various real-world occlusion conditions, dubbed Occlu-FER. Extensive experiments on both public benchmarks and Occlu-FER demonstrate that our proposed ORSANet achieves SOTA recognition performance. Code is publicly available at https://github.com/Wenyuzhy/ORSANet-master.

Figures

Figures reproduced from arXiv: 2507.15401 by the authors.

Figure 1
Figure 1. Motivation comparison. Existing methods (top) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Our main contributions are as follows: • New Method. We propose ORSANet, which introduces both dense semantics prior (i.e., semantic segmentation maps) and sparse geometric prior (i.e., facial landmarks) as explicit semantic guidance to learn high-level semantic knowledge to deal with the challenging occlusion issue, achieving SOTA performance on several widely used benchmarks. • New Interaction Mechanism. We custom… view at source ↗
Figure 2
Figure 2. Samples in our Occlu-FER dataset. The last two rows [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: Illustration of ORSANet. (a) illustrates the overall pipeline. The sparse geometric prior and dense semantics prior are [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 4
Figure 4. Figure 4: Visualization of facial expression. Including input [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Visual examples of deficiencies. When facing low [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

76 extracted references · 63 canonical work pages

  1. [1]

    Faiza Abdat, Choubeila Maaoui, and Alain Pruski. 2011. Human-computer in- teraction using emotion recognition from facial expression. In2011 UKSim 5th European Symposium on Computer Modeling and Simulation. 196–201

    work page 2011

  2. [2]

    Xinran Cao, Liang Luo, Yu Gu, and Fuji Ren. 2025. Co-dance with Ambigu- ity: An Ambiguity-Aware Facial Expression Recognition Framework for More Robustness.IEEE Transactions on Circuits and Systems for Video Technology (2025)

    work page 2025

  3. [3]

    Cunjian Chen. 2021. PyTorch Face Landmark: A fast and accurate facial landmark detector.Opensource software available at https://github. com/cunjian/pytorchf acelandmark(2021), 27

    work page 2021

  4. [4]

    Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. 2021. Crossvit: Cross- attention multi-scale vision transformer for image classification. InProceedings of the IEEE/CVF international conference on computer vision. 357–366

    2021

  5. [5]

    Tianshui Chen, Tao Pu, Hefeng Wu, Yuan Xie, Lingbo Liu, and Liang Lin. 2021. Cross-domain facial expression recognition: A unified evaluation benchmark and adversarial graph learning.IEEE transactions on pattern analysis and machine intelligence44 (2021), 9887–9903

    work page 2021

  6. [6]

    M Kalpana Chowdary, Tu N Nguyen, and D Jude Hemanth. 2023. Deep learning- based facial emotion recognition for human–computer interaction applications. Neural Computing and Applications35 (2023), 23311–23328

    work page 2023

  7. [7]

    Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. Arcface: Additive angular margin loss for deep face recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4690–4699

    2019

  8. [8]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    arXiv 2020

Show all 76 references
  1. [9]

    Huilin Ge, Zhiyu Zhu, Yuewei Dai, Biao Wang, and Xuedong Wu. 2022. Facial expression recognition based on deep learning.Computer Methods and Programs in Biomedicine215 (2022), 106621

    2022

  2. [10]

    Jiuxiang Gu, Zhenhua Wang, Jason Kuen, Lianyang Ma, Amir Shahroudy, Bing Shuai, Ting Liu, Xingxing Wang, Gang Wang, Jianfei Cai, et al . 2018. Recent advances in convolutional neural networks.Pattern recognition77 (2018), 354– 377

    2018

  3. [11]

    Tianchu Guo, Youngsung Kim, Hui Zhang, Deheng Qian, ByungIn Yoo, Jingtao Xu, Dongqing Zou, Jae-Joon Han, and Changkyu Choi. 2018. Residual encoder decoder network and adaptive prior for face parsing. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32

    2018

  4. [12]

    Yandong Guo, Lei Zhang, Yuxiao Hu, Xiaodong He, and Jianfeng Gao. 2016. Ms- celeb-1m: A dataset and benchmark for large-scale face recognition. InEuropean Conference on Computer Vision. 87–102

    2016

  5. [13]

    Danfeng Hong, Zhu Han, Jing Yao, Lianru Gao, Bing Zhang, Antonio Plaza, and Jocelyn Chanussot. 2021. SpectralFormer: Rethinking hyperspectral image classification with transformers.IEEE Transactions on Geoscience and Remote Sensing60 (2021), 1–15

    2021

  6. [14]

    Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger

  7. [15]

    Haibo Jin, Shengcai Liao, and Ling Shao. 2021. Pixel-in-pixel net: Towards efficient facial landmark detection in the wild.International Journal of Computer Vision 129 (2021), 3174–3194

    2021

  8. [16]

    Peiguang Jing, Xianyi Liu, Ji Wang, Yinwei Wei, Liqiang Nie, and Yuting Su

  9. [17]

    Lingzhao Ju and Xu Zhao. 2022. Mask-based attention parallel network for in-the-wild facial expression recognition. InIEEE International Conference on Acoustics, Speech and Signal Processing. 2410–2414

    2022

  10. [18]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization.arXiv preprint arXiv:1412.6980(2014)

    2014 arXiv

  11. [19]

    Diederik P Kingma and Max Welling. 2013. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114(2013)

    2013 arXiv

  12. [20]

    Thomas Kopalidis, Vassilios Solachidis, Nicholas Vretos, and Petros Daras. 2024. Advances in facial expression recognition: A survey of methods, benchmarks, models, and datasets.Information15 (2024), 135

    2024

  13. [21]

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

    2009

  14. [22]

    Naveen Kumar HN, A Suresh Kumar, Guru Prasad MS, and Mohd Asif Shah. 2023. Automatic facial expression recognition combining texture and shape features from prominent facial regions.IET Image Processing17 (2023), 1111–1125

    2023

  15. [23]

    Isack Lee, Eungi Lee, and Seok Bong Yoo. 2023. Latent-ofer: Detect, mask, and reconstruct with latent vectors for occluded facial expression recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision. 1536– 1546

    2023

  16. [24]

    Hanting Li, Hongjing Niu, Zhaoqing Zhu, and Feng Zhao. 2024. Cliper: A unified vision-language framework for in-the-wild facial expression recognition. InIEEE International Conference on Multimedia and Expo. 1–6

    2024

  17. [25]

    Shan Li and Weihong Deng. 2020. Deep facial expression recognition: A survey. IEEE transactions on affective computing13 (2020), 1195–1215

    2020

  18. [26]

    Shan Li, Weihong Deng, and JunPing Du. 2017. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. InProceedings of the IEEE conference on computer vision and pattern recognition. 2852–2861

    2017

  19. [27]

    Xin Li, Shangfei Wang, and Xuandong Huang. 2024. Temporal Enhancement for Video Affective Content Analysis. InProceedings of the 32nd ACM International Conference on Multimedia. 642–650

    2024

  20. [28]

    Yande Li, Mingjie Wang, Minglun Gong, Yonggang Lu, and Li Liu. 2024. FER- former: Multimodal Transformer for Facial Expression Recognition.IEEE Trans- actions on Multimedia(2024)

    2024

  21. [29]

    Tsung-Yi Lin, Piotr Dollár, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. 2017. Feature pyramid networks for object detection. InProceed- ings of the IEEE conference on computer vision and pattern recognition. 2117–2125

    2017

  22. [30]

    Yiming Lin, Jie Shen, Yujiang Wang, and Maja Pantic. 2021. Roi tanh-polar transformer network for face parsing in the wild.Image and Vision Computing 112 (2021), 104190

    2021

  23. [31]

    Hanwei Liu, Huiling Cai, Qingcheng Lin, Xuefeng Li, and Hui Xiao. 2023. Learn- ing from more: Combating uncertainty cross-multidomain for facial expression recognition. InProceedings of the 31st ACM International Conference on Multime- dia. 5889–5898

    2023

  24. [32]

    Hai Liu, Qiyun Zhou, Cheng Zhang, Junyan Zhu, Tingting Liu, Zhaoli Zhang, and You-Fu Li. 2024. MMATrans: Muscle movement aware representation learning for facial expression recognition via transformers.IEEE Transactions on Industrial Informatics(2024)

    2024

  25. [33]

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. 2015. Fully convolutional networks for semantic segmentation. InProceedings of the IEEE conference on computer vision and pattern recognition. 3431–3440

    2015

  26. [34]

    Yanzuo Lu, Manlin Zhang, Andy J Ma, Xiaohua Xie, and Jianhuang Lai. 2024. Coarse-to-fine latent diffusion for pose-guided person image synthesis. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 6420–6429

    2024

  27. [35]

    Jiawei Mao, Rui Xu, Xuesong Yin, Yuanqi Chang, Binling Nie, Aibin Huang, and Yigang Wang. 2024. Poster++: A simpler and stronger facial expression recognition network.Pattern Recognition(2024), 110951

    2024

  28. [36]

    Ali Mollahosseini, Behzad Hasani, and Mohammad H Mahoor. 2017. Affectnet: A database for facial expression, valence, and arousal computing in the wild.IEEE Transactions on Affective Computing10 (2017), 18–31

    2017

  29. [37]

    Kartik Narayan, Vibashan VS, and Vishal M Patel. 2024. Segface: Face segmenta- tion of long-tail classes.arXiv preprint arXiv:2412.08647(2024)

    2024 arXiv

  30. [38]

    Bowen Pan, Shangfei Wang, and Bin Xia. 2019. Occluded facial expression recognition enhanced through privileged information. InProceedings of the 27th ACM international conference on multimedia. 566–573

    2019

  31. [39]

    Jicai Pan and Shangfei Wang. 2023. Progressive visual content understanding net- work for image emotion classification. InProceedings of the 31st ACM International Conference on Multimedia. 6034–6044

    2023

  32. [40]

    Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. 2019. Semantic image synthesis with spatially-adaptive normalization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2337–2346

    2019

  33. [41]

    Yongming Rao, Wenliang Zhao, Zheng Zhu, Jiwen Lu, and Jie Zhou. 2021. Global filter networks for image classification.Advances in neural information processing systems34 (2021), 980–993

    2021

  34. [42]

    I Michael Revina and WR Sam Emmanuel. 2021. A survey on human face expression recognition techniques.Journal of King Saud University-Computer and Information Sciences33 (2021), 619–628

    2021

  35. [43]

    Delian Ruan, Yan Yan, Shenqi Lai, Zhenhua Chai, Chunhua Shen, and Hanzi Wang

  36. [44]

    Muhammad Sajjad, Adnan Shah, Zahoor Jan, Syed Inayat Shah, Sung Wook Baik, and Irfan Mehmood. 2018. Facial appearance and texture feature-based robust facial expression recognition framework for sentiment knowledge discovery. Cluster Computing21 (2018), 549–567

    2018

  37. [45]

    MM ’25, October 27–31, 2025, Dublin, Ireland

    Muhammad Sajjad, Fath U Min Ullah, Mohib Ullah, Georgia Christodoulou, Faouzi Alaya Cheikh, Mohammad Hijji, Khan Muhammad, and Joel JPC Rodrigues. MM ’25, October 27–31, 2025, Dublin, Ireland. Huiyu Zhai, Xingxing Yang, Yalan Ye, Chenyang Li, Bin Fan, Changze Li

    2025

  38. [46]

    Dinh Viet Sang, Pham Thai Ha, et al. 2018. Discriminative deep feature learning for facial emotion recognition. In1st International Conference on Multimedia Analysis and Pattern Recognition. 1–6

    2018

  39. [47]

    Andrey V Savchenko. 2021. Facial expression and attributes recognition based on multi-task learning of lightweight neural networks. InIEEE 19th international symposium on intelligent systems and informatics. 119–124

    2021

  40. [48]

    Jiahui She, Yibo Hu, Hailin Shi, Jun Wang, Qiu Shen, and Tao Mei. 2021. Dive into ambiguity: Latent distribution mining and pairwise uncertainty estimation for facial expression recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6248–6257

    2021

  41. [49]

    A comprehensive survey on deep facial expression recognition: challenges, applications, and future guidelines.Alexandria Engineering Journal68 (2023), 817–840

    2023

  42. [50]

    Gusi Te, Wei Hu, Yinglu Liu, Hailin Shi, and Tao Mei. 2021. Agrnet: Adaptive graph representation learning and reasoning for face parsing.IEEE Transactions on Image Processing30 (2021), 8236–8250

    2021

  43. [51]

    Gusi Te, Yinglu Liu, Wei Hu, Hailin Shi, and Tao Mei. 2020. Edge-aware graph representation learning and reasoning for face parsing. InEuropean Conference on Computer Vision. 258–274

    2020

  44. [52]

    Thanh-Hung Vo, Guee-Sang Lee, Hyung-Jeong Yang, and Soo-Hyung Kim. 2020. Pyramid with super resolution for in-the-wild facial expression recognition.IEEE Access8 (2020), 131988–132001

    2020

  45. [53]

    Zeng Tao, Yan Wang, Zhaoyu Chen, Boyang Wang, Shaoqi Yan, Kaixun Jiang, Shuyong Gao, and Wenqiang Zhang. 2023. Freq-hd: An interpretable frequency- based high-dynamics affective clip selection method for in-the-wild facial expres- sion recognition in videos. InProceedings of t...

    2023

  46. [54]

    Kai Wang, Xiaojiang Peng, Jianfei Yang, Shijian Lu, and Yu Qiao. 2020. Suppress- ing uncertainties for large-scale facial expression recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 6897–6906

    2020

  47. [55]

    Kai Wang, Xiaojiang Peng, Jianfei Yang, Debin Meng, and Yu Qiao. 2020. Region attention networks for pose and occlusion robust facial expression recognition. IEEE Transactions on Image Processing29 (2020), 4057–4069

    2020

  48. [56]

    Qingzhong Wang, Pengfei Zhang, Haoyi Xiong, and Jian Zhao. 2021. Face. evolve: A high-performance face recognition library.arXiv preprint arXiv:2107.08621 (2021)

    2021 arXiv

  49. [57]

    Jingdong Wang, Ke Sun, Tianheng Cheng, Borui Jiang, Chaorui Deng, Yang Zhao, Dong Liu, Yadong Mu, Mingkui Tan, Xinggang Wang, et al. 2020. Deep high-resolution representation learning for visual recognition.IEEE transactions on pattern analysis and machine intelligence43 (2020...

    2020

  50. [58]

    Chang Xu, Jian Ding, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. 2023. Dynamic coarse-to-fine learning for oriented tiny object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7318–7328

    2023

  51. [59]

    Fanglei Xue, Qiangchang Wang, and Guodong Guo. 2021. Transfer: Learning relation-aware facial expression representations with transformers. InProceedings of the IEEE/CVF International conference on computer vision. 3601–3610

    2021

  52. [60]

    Dingkang Yang, Kun Yang, Mingcheng Li, Shunli Wang, Shuaibing Wang, and Lihua Zhang. 2024. Robust emotion recognition in context debiasing. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 12447–12457

    2024

  53. [61]

    Daiqing Wu, Dongbao Yang, Yu Zhou, and Can Ma. 2024. Bridging Visual Affective Gap: Borrowing Textual Knowledge by Learning from Noisy Image- Text Pairs. InProceedings of the 32nd ACM International Conference on Multimedia. 602–611

    2024

  54. [62]

    Xingxing Yang, Jie Chen, and Zaifeng Yang. 2024. Hyperspectral image recon- struction via combinatorial embedding of cross-channel spatio-spectral clues. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38. 6567–6575

    2024

  55. [63]

    Jizheng Yi, Xia Mao, Lijiang Chen, Yuli Xue, and Angelo Compare. 2014. Facial expression recognition considering individual differences in facial structure and texture.IET Computer Vision8 (2014), 429–440

    2014

  56. [64]

    Jun Yu, Zhihong Wei, Zhongpeng Cai, Gongpeng Zhao, Zerui Zhang, Yongqi Wang, Guochen Xie, Jichao Zhu, Wangyuan Zhu, Qingsong Liu, et al . 2024. Exploring Facial Expression Recognition through Semi-Supervised Pre-training and Temporal Modeling. InProceedings of the IEEE/CVF Con...

    2024

  57. [65]

    Xingxing Yang, Jie Chen, and Zaifeng Yang. 2023. Cooperative colorization: Exploring latent cross-domain priors for nir image spectrum translation. In Proceedings of the 31st ACM International Conference on Multimedia. 2409–2417

    2023

  58. [66]

    Yuhang Zhang, Chengrui Wang, Xu Ling, and Weihong Deng. 2022. Learn from all: Erasing attention consistency for noisy label facial expression recognition. In European Conference on Computer Vision. 418–434

    2022

  59. [67]

    Jianing Zhao, Jingjing Wang, Yujie Jin, Jiamin Luo, and Guodong Zhou. 2024. Hawkeye: Discovering and Grounding Implicit Anomalous Sentiment in Recon- videos via Scene-enhanced Video Large Language Model. InProceedings of the 32nd ACM International Conference on Multimedia. 592–601

    2024

  60. [68]

    Zengqun Zhao, Qingshan Liu, and Feng Zhou. 2021. Robust lightweight facial expression recognition network with label distribution training. InProceedings of the AAAI conference on artificial intelligence, Vol. 35. 3510–3519

    2021

  61. [69]

    Dan Zeng, Zhiyuan Lin, Xiao Yan, Yuting Liu, Fei Wang, and Bo Tang. 2022. Face2exp: Combating data biases for facial expression recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 20291– 20300

    2022

  62. [70]

    Qingping Zheng, Jiankang Deng, Zheng Zhu, Ying Li, and Stefanos Zafeiriou. 2022. Decoupled multi-task learning with cyclical self-regulation for face parsing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 4156–4165

    2022

  63. [71]

    Yisu Zhou, Xiaolin Hu, and Bo Zhang. 2015. Interlinked convolutional neural networks for face parsing. In12th International Symposium on Neural Networks. 222–231

    2015

  64. [72]

    Junjie Zhu, Bingjun Luo, Ao Sun, Jinghang Tan, Xibin Zhao, and Yue Gao. 2023. Variance-Aware Bi-Attention Expression Transformer for Open-Set Facial Ex- pression Recognition in the Wild. InProceedings of the 31st ACM International Conference on Multimedia. 862–870

    2023

  65. [73]

    Ce Zheng, Matias Mendieta, and Chen Chen. 2023. Poster: A pyramid cross- fusion transformer network for facial expression recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision. 3146–3155

    2023

  66. [2017]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    Densely connected convolutional networks. InProceedings of the IEEE conference on computer vision and pattern recognition. 4700–4708

  67. [2021]

    InProceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Feature decomposition and reconstruction learning for effective facial expression recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 7660–7669

  68. [2023]

    InProceedings of the 31st ACM International Conference on Multimedia

    StyleEDL: Style-Guided High-order Attention Network for Image Emotion Distribution Learning. InProceedings of the 31st ACM International Conference on Multimedia. 853–861

Pith tools

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