Pith. sign in

REVIEW 4 major objections 6 minor 34 references

MERba: Multi-Receptive Field MambaVision for Micro-Expression Recognition

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read MERba, a hybrid MambaVision architecture, reports state-of-the-art micro-expression recognition on three benchmarks while using far fewer parameters than the leading Transformer baseline.

desk verdict A sensible architecture and strong parameter efficiency are undermined by an overblown SOTA claim and missing statistical rigor, but the paper deserves peer review. read the letter →

arxiv 2506.14468 v2 pith:J7SM6P4U submitted 2025-06-17 cs.HC

Xinglong Mao , Shifeng Liu , Sirui Zhao , Tong Xu , Hanchao Wang , Baozhi Jia , Enhong Chen This is my paper · ORCID
classification cs.HC
keywords micro-expressionrecognitionMambaVisionstatespacemodelsmulti-receptivefieldasymmetricmulti-scanningdual-granularityclassificationlocal-globalfeatureintegrationopticalflow
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

This paper proposes MERba, an architecture for micro-expression recognition that combines local window-based MambaVision mixers with global self-attention. It claims the design captures both subtle localized muscle movements and whole-face dependencies, which prior CNN and Transformer methods handle only separately. Across three benchmark datasets, MERba reports the best published results: UF1 of 0.8840 on 3DB-Combined, UF1 of 0.4332 on DFME-public Test A, and 75.2% accuracy on MMEW, using about 101 million parameters while the strongest prior model uses 438 million. A dual-granularity classification head is added for seven-class datasets, where negative emotions such as anger and disgust are easily confused.

What carries the argument

The load-bearing object is the Local-Global Feature Integration (LGFI) stage, which pairs a MERba Local Extractor with global self-attention. Within each local window, the extractor rearranges pixels into four 1D sequences using the asymmetric multi-scanning strategy, passes each through a MambaVision Mixer, reconstructs and sums the four feature maps, and then refines them with a linear layer and MLP. Two multi-head self-attention blocks then model inter-window dependencies. The Dual-Granularity Classification Module (DGCM) is the second mechanism, decomposing classification into coarse polarity and fine negative-emotion heads with a progressively weighted loss.

What would settle it

A concrete test: train MERba on DFME-public with the four asymmetric scan directions replaced by a bi-directional or symmetric set and compare UF1; if UF1 does not drop, the redundancy argument fails. A second test on unaligned faces would also settle it, since the symmetry-redundancy claim predicts that asymmetric scanning is no better than symmetric scanning only after alignment.

Watch

Extended reading notes

Core claim

The central claim is that micro-expression recognition is best served by a hierarchy of local-global integration stages: in each stage, the face is divided into non-overlapping 7x7 windows, and within each window a MambaVision mixer with four asymmetric scanning directions (horizontal raster, vertical raster, horizontal zigzag, vertical zigzag) extracts fine motion patterns, after which lightweight self-attention layers aggregate the windows into a global representation. Progressive downsampling across stages widens the receptive field until the final stage covers the whole face. The paper also introduces a Dual-Granularity Classification Module that trains a coarse polarity head and, only for negative samples, a fine-grained emotion head, reporting that this reduces confusion among anger, disgust, fear, and sadness. The paper argues that asymmetric scans avoid the redundancy of bi-directional or symmetric scans because aligned faces already have left-right symmetry.

Load-bearing premise

The load-bearing premise is that four fixed asymmetric scanning directions—horizontal and vertical raster, horizontal and vertical zigzag—capture all the local spatial dependencies a micro-expression recognizer needs, while bi-directional and symmetric scans are purely redundant.

Editorial extensions

If this is right

  • Micro-expression recognition can reach state-of-the-art results with about 101 million parameters, roughly a quarter of HTNet's 438 million, so parameter-heavy hierarchical Transformers are not required.
  • For aligned faces, bi-directional and symmetric scans add redundancy; future Mamba-based face models can likely use a smaller asymmetric scan set without losing accuracy.
  • Separating emotion polarity from fine-grained category prediction is a workable remedy for inter-class confusion among negative micro-expressions.
  • Combining local window extraction with global attention is necessary: local-only or global-only versions of the model score at least 4 percentage points lower in UF1 on DFME-public Test B.

Reading between the lines

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

  • Editorial inference: the paper leaves scan-direction choice as a fixed empirical choice; a learned per-window or per-stage scan-direction search could test whether exactly four directions are optimal.
  • Editorial inference: the symmetry-redundancy rationale depends on aligned, front-facing faces; if unaligned or profile faces are common, symmetric scans may stop being redundant and the advantage could invert.
  • Editorial inference: the dual-granularity loss only supervises the fine head on negative samples, so the DGCM benefit may transfer to other fine-grained classification tasks where a coarse polarity split separates the hard classes.
  • Editorial inference: on MMEW, MERba ties MA2MI without macro-expression pretraining; this suggests task-adapted architecture can substitute for large-scale pretraining in small-sample affective recognition, a claim worth testing on other small affective datasets.
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. This manuscript proposes MERba, a hierarchical MambaVision-based architecture for micro-expression recognition. The model combines local windowed MambaVision mixers with global self-attention in three Local-Global Feature Integration stages, uses an asymmetric four-direction scanning strategy, and adds a Dual-Granularity Classification Module (DGCM) to separate coarse emotion polarity from fine-grained negative-emotion categories. Experiments on 3DB-Combined, DFME-public, and MMEW report UF1/ACC numbers that are state-of-the-art or competitive, together with ablations and Grad-CAM visualizations.

Significance. If the reported results are reproducible, MERba is a meaningful contribution: it is substantially smaller than HTNet (101.21M vs 438.51M parameters) while reaching competitive accuracy, and the DGCM coarse-to-fine loss is a simple, plausibly effective idea for class-imbalanced micro-expression benchmarks. The paper includes component ablations, comparison to a reproduced MambaVision baseline, and visual case studies, which strengthen the empirical narrative. The main weakness is that the central claim of state-of-the-art performance rests on unreplicated point estimates and a partially underspecified evaluation protocol.

major comments (4)
  1. [Abstract and §IV-C] The abstract and conclusion claim that MERba 'consistently outperforms existing methods,' but Tables IV and VI contradict this: on CASME II, MERba's UF1 (0.9498) is lower than HTNet's 0.9532, and on MMEW, MERba-DGCM ties MA2MI at 75.2% accuracy. Please qualify the central claim to 'competitive or state-of-the-art on most benchmarks' and discuss the specific exceptions in the text.
  2. [§IV-B and Tables IV–VI] No standard deviations, repeated-seed runs, or significance tests are reported anywhere. Given the small dataset sizes (N=442, 234, and 474 for the main test sets), margins of 1–3 points (e.g., 0.8840 vs 0.8603 UF1 on 3DB-Combined) are within plausible run-to-run noise. Please report mean±std over at least three random seeds, or fold-wise results, and a paired significance test for the headline comparisons.
  3. [§IV-B] Early stopping is mentioned but the validation split is not defined. In the LOSO protocol for 3DB-Combined and the fixed train/test protocol for DFME, early stopping on the held-out test subjects or test set would leak information and inflate results. Please specify the validation procedure (e.g., an inner split of the training subjects, or a fixed held-out validation set) and report the epoch at which training stopped for each protocol.
  4. [§III-B and Table VIII] The choice of the four scanning directions is justified only by the ablation. While Table VIII shows that a+b+c+d outperforms the tested alternatives, the paper does not establish why these four directions, rather than other asymmetric sets, are optimal, nor does it test the claim that bi-directional/symmetric scans are redundant for micro-expressions. Providing a quantitative coverage analysis of token adjacencies or an ablation across additional direction sets would make this load-bearing design choice more robust.
minor comments (6)
  1. [Table I] The 'Global Self-attention 2' entries are unclear; please specify whether two MHSA blocks are used in every LGFI stage and align the row/column formatting so the block counts are unambiguous.
  2. [§III-B] The statement that motion features from symmetrical left and right facial regions are 'identical' is not obviously true, because micro-expressions can be asymmetric; please soften or support this claim with evidence.
  3. [Tables IV–VI] The footnotes say MambaVision-B is 'reproduced with the same parameters as MERba'; please clarify whether this means identical hyperparameters or identical model size, and report MambaVision-B's parameter count for a fair efficiency comparison.
  4. [§IV-B] No code or configuration files are provided; releasing the implementation and the exact data preprocessing pipeline would materially aid reproducibility, especially given the early-stopping ambiguity.
  5. [Figure 2] The four scanning directions are hard to distinguish in grayscale; please use distinct colors or labels directly on the paths to make the asymmetric patterns visible.
  6. [Reference [31]] Reference [31] is cited as a URL; please include a version/access date and, if available, a peer-reviewed description of the CCAC 2024 challenge protocol.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the SOTA claim is an empirical benchmark comparison, not a derivation that reduces to its inputs.

full rationale

The paper's central claim is empirical state-of-the-art performance on three external benchmark datasets (3DB-Combined, DFME-public, MMEW). MERba is a proposed architecture, and its components (Local-Global Feature Integration stages, asymmetric multi-scanning, Dual-Granularity Classification Module) are evaluated through held-out benchmark comparisons and ablations. None of the reported results are defined in terms of the model's own fitted constants: the loss in Eq. (6), the scanning fusion in Eq. (3), and the global aggregation in Eq. (5) are forward operations whose outputs are measured against external labels, not reconstructed from those labels. The self-citations present in the paper—DFME dataset [30], ME-PLAN apex detection [23], and the CCAC leaderboard [31]—are used as data sources, preprocessing tools, or external baseline references; they do not supply the conclusion that MERba outperforms prior methods. No self-citation chain forces the reported numbers, no uniqueness theorem is imported from the authors' prior work, and no known result is merely renamed. Concerns about single-run results, absent error bars, and an unspecified early-stopping split are reproducibility or correctness risks, not circularity, and per the analysis rules they do not raise the circularity score.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

MERba does not introduce new physical entities or forces; its novelty lies in architectural design and training strategy. The free parameters listed are hyperparameters chosen by the authors, not derived from data fitting. The axioms are domain assumptions about the input representation, scanning strategy, and classification hierarchy.

free parameters (4)
  • Window size = 7x7
    Window size for local scanning is fixed at 7x7 across all LGFI stages, chosen empirically without justification.
  • Number of scanning directions = 4
    Four asymmetric directions are used, selected based on the authors' reasoning about facial symmetry and redundancy.
  • Alpha schedule for DGCM loss = min(0.5 + 2.0 * epoch/total_epochs, 2.0)
    The weighting factor alpha is hand-designed to increase during training; its specific form is not derived from first principles.
  • Block depths (N1, N2, N3, N4) = 3, 2, 6, 4
    Stage depths are architectural hyperparameters chosen by the authors; they are not tuned via a validation set in the paper.
assumptions (3)
  • domain assumption Optical flow (TV-L1) computed from onset and apex frames is a sufficient representation of micro-expression motion.
    The model input is fixed as TV-L1 optical flow, a standard choice in MER, but the paper does not compare with raw frames or other motion representations.
  • ad hoc to paper The four scanning directions (raster horizontal, raster vertical, zigzag horizontal, zigzag vertical) capture necessary spatial dependencies while avoiding redundancy.
    This is a core design assumption; the paper provides empirical ablation but no theoretical justification for the optimality of these four directions.
  • domain assumption The coarse-to-fine classification scheme, where the fine head is only active for negative emotions, improves discriminability because negative emotions are most confusable.
    The paper assumes that inter-class confusion is highest among negative emotions and focuses the fine head there, but it does not quantitatively validate this assumption across all datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MERba: Multi-Receptive Field MambaVision for Micro-Expression Recognition." pith.science (2026). https://pith.science/paper/J7SM6P4U

@misc{pith2026250614468,
  author       = {Pith},
  title        = {Pith review of: MERba: Multi-Receptive Field MambaVision for Micro-Expression Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7SM6P4U}},
  note         = {Machine review of arXiv:2506.14468}
}
read the original abstract

Micro-expressions (MEs) are brief, involuntary facial movements that reveal genuine emotions, offering valuable insights for psychological assessment and criminal investigations. Despite significant progress in automatic ME recognition (MER), existing methods still struggle to simultaneously capture localized muscle activations and global facial dependencies, both essential for decoding subtle emotional cues. To address this challenge, we propose MERba, a hierarchical multi-receptive field architecture specially designed for MER, which incorporates a series of Local-Global Feature Integration stages. Within each stage, detailed intra-window motion patterns are captured using MERba Local Extractors, which integrate MambaVision Mixers with a tailored asymmetric multi-scanning strategy to enhance local spatial sensitivity. These localized features are then aggregated through lightweight self-attention layers that explicitly model inter-window relationships, enabling effective global context construction. Furthermore, to mitigate the challenge of high inter-class similarity among negative MEs, we introduce a Dual-Granularity Classification Module that decomposes the recognition task into a coarse-to-fine paradigm. Extensive experiments on three benchmark datasets demonstrate that MERba consistently outperforms existing methods, with ablation studies confirming the effectiveness of each proposed component.

Figures

Figures reproduced from arXiv: 2506.14468 by the authors.

Figure 1
Figure 1. The overall framework of MERba, with the specific structures [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of Asymmetric Multi-Scanning Strategy. The four [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Grad-CAM visualizations of ME samples from DFME. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 30 canonical work pages

  1. [1]

    Video-based facial micro-expression analysis: A survey of datasets, features and algorithms,

    Xianye Ben, Yi Ren, Junping Zhang, Su-Jing Wang, Kidiyo Kpalma, Weixiao Meng, and Yong-Jin Liu, “Video-based facial micro-expression analysis: A survey of datasets, features and algorithms,” IEEE transac- tions on pattern analysis and machine intelligence , vol. 44, no. 9, pp. 5826–5846, 2021

    work page 2021

  2. [2]

    How fast are the leaked facial expressions: The duration of micro- expressions,

    Wen-Jing Yan, Qi Wu, Jing Liang, Yu-Hsin Chen, and Xiaolan Fu, “How fast are the leaked facial expressions: The duration of micro- expressions,” Journal of Nonverbal Behavior , vol. 37, pp. 217–230, 2013

    work page 2013

  3. [3]

    Paul Ekman, Telling lies: Clues to deceit in the marketplace, politics, and marriage (revised edition) , WW Norton & Company, 2009

    work page 2009

  4. [4]

    Elderly depression recognition based on facial micro- expression extraction.,

    Wei Huang, “Elderly depression recognition based on facial micro- expression extraction.,” Traitement du Signal , vol. 38, no. 4, 2021

    work page 2021

  5. [5]

    Shallow triple stream three-dimensional cnn (ststnet) for micro-expression recognition,

    Sze-Teng Liong, Yee Siang Gan, John See, Huai-Qian Khor, and Yen- Chang Huang, “Shallow triple stream three-dimensional cnn (ststnet) for micro-expression recognition,” in 2019 14th IEEE international conference on automatic face & gesture recognition (FG 2019) . IEEE, 2019, pp. 1–5

    work page 2019

  6. [6]

    A two-stage 3d cnn based learning method for spontaneous micro-expression recognition,

    Sirui Zhao, Hanqing Tao, Yangsong Zhang, Tong Xu, Kun Zhang, Zhongkai Hao, and Enhong Chen, “A two-stage 3d cnn based learning method for spontaneous micro-expression recognition,” Neurocomput- ing, vol. 448, pp. 276–289, 2021

    work page 2021

  7. [7]

    Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition,

    Ling Zhou, Qirong Mao, Xiaohua Huang, Feifei Zhang, and Zhihong Zhang, “Feature refinement: An expression-specific feature learning and fusion method for micro-expression recognition,” Pattern Recognition, vol. 122, pp. 108275, 2022

    work page 2022

  8. [8]

    A novel magnification-robust network with sparse self- attention for micro-expression recognition,

    Mengting Wei, Wenming Zheng, Xingxun Jiang, Yuan Zong, Cheng Lu, and Jiateng Liu, “A novel magnification-robust network with sparse self- attention for micro-expression recognition,” in 2022 26th International Conference on Pattern Recognition (ICPR). IEEE, 2022, pp. 1120–1126

    work page 2022

Show all 34 references
  1. [9]

    Mfdan: Multi-level flow-driven attention network for micro-expression recognition,

    Wenhao Cai, Junli Zhao, Ran Yi, Minjing Yu, Fuqing Duan, Zhenkuan Pan, and Yong-Jin Liu, “Mfdan: Multi-level flow-driven attention network for micro-expression recognition,” IEEE Transactions on Circuits and Systems for Video Technology , 2024

    2024

  2. [10]

    Htnet for micro-expression recognition,

    Zhifeng Wang, Kaihao Zhang, Wenhan Luo, and Ramesh Sankara- narayana, “Htnet for micro-expression recognition,” Neurocomputing, vol. 602, pp. 128196, 2024

    2024

  3. [11]

    Mamba: Linear-time sequence modeling with selective state spaces,

    Albert Gu and Tri Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” arXiv preprint arXiv:2312.00752 , 2023

    2023 arXiv

  4. [12]

    A new approach to linear filtering and prediction problems,

    Rudolph Emil Kalman, “A new approach to linear filtering and prediction problems,” 1960

    1960

  5. [13]

    Vision mamba: Efficient visual represen- tation learning with bidirectional state space model,

    Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang, “Vision mamba: Efficient visual represen- tation learning with bidirectional state space model,” arXiv preprint arXiv:2401.09417, 2024

    2024 arXiv

  6. [14]

    Vmamba: Visual state space model,

    Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, Jianbin Jiao, and Yunfan Liu, “Vmamba: Visual state space model,” Advances in neural information processing systems , vol. 37, pp. 103031–103063, 2024

    2024

  7. [15]

    Mambavision: A hybrid mamba- transformer vision backbone,

    Ali Hatamizadeh and Jan Kautz, “Mambavision: A hybrid mamba- transformer vision backbone,” in Proceedings of the Computer Vision and Pattern Recognition Conference , 2025, pp. 25261–25270

    2025

  8. [16]

    Facial action coding system,

    Paul Ekman and Wallace V Friesen, “Facial action coding system,” Environmental Psychology & Nonverbal Behavior , 1978

    1978

  9. [17]

    Localmamba: Visual state space model with windowed selective scan,

    Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu, “Localmamba: Visual state space model with windowed selective scan,” in European Conference on Computer Vision. Springer, 2025, pp. 12–22

    2025

  10. [18]

    Plainmamba: Improving non-hierarchical mamba in visual recognition,

    Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J Crowley, “Plainmamba: Improving non-hierarchical mamba in visual recognition,”arXiv preprint arXiv:2403.17695, 2024

    2024 arXiv

  11. [19]

    Dynamic texture recognition using local binary patterns with an application to facial expressions,

    Guoying Zhao and Matti Pietikainen, “Dynamic texture recognition using local binary patterns with an application to facial expressions,” IEEE transactions on pattern analysis and machine intelligence , vol. 29, no. 6, pp. 915–928, 2007

    2007

  12. [20]

    Less is more: Micro-expression recognition from video using apex frame,

    Sze-Teng Liong, John See, KokSheik Wong, and Raphael C-W Phan, “Less is more: Micro-expression recognition from video using apex frame,” Signal Processing: Image Communication , vol. 62, pp. 82–92, 2018

    2018

  13. [21]

    A main directional mean optical flow feature for spontaneous micro-expression recognition,

    Yong-Jin Liu, Jin-Kai Zhang, Wen-Jing Yan, Su-Jing Wang, Guoying Zhao, and Xiaolan Fu, “A main directional mean optical flow feature for spontaneous micro-expression recognition,” IEEE Transactions on Affective Computing, vol. 7, no. 4, pp. 299–310, 2015

    2015

  14. [22]

    Micro-expression recog- nition with small sample size by transferring long-term convolutional neural network,

    Su-Jing Wang, Bing-Jun Li, Yong-Jin Liu, Wen-Jing Yan, Xinyu Ou, Xiaohua Huang, Feng Xu, and Xiaolan Fu, “Micro-expression recog- nition with small sample size by transferring long-term convolutional neural network,” Neurocomputing, vol. 312, pp. 251–262, 2018

    2018

  15. [23]

    Me-plan: A deep prototyp- ical learning with local attention network for dynamic micro-expression recognition,

    Sirui Zhao, Huaying Tang, Shifeng Liu, Yangsong Zhang, Hao Wang, Tong Xu, Enhong Chen, and Cuntai Guan, “Me-plan: A deep prototyp- ical learning with local attention network for dynamic micro-expression recognition,” Neural networks, vol. 153, pp. 427–443, 2022

    2022

  16. [24]

    Short and long range relation based spatio-temporal transformer for micro-expression recognition,

    Liangfei Zhang, Xiaopeng Hong, Ognjen Arandjelovi ´c, and Guoying Zhao, “Short and long range relation based spatio-temporal transformer for micro-expression recognition,” IEEE Transactions on Affective Computing, vol. 13, no. 4, pp. 1973–1985, 2022

    1973

  17. [25]

    A duality based approach for realtime tv-l 1 optical flow,

    Christopher Zach, Thomas Pock, and Horst Bischof, “A duality based approach for realtime tv-l 1 optical flow,” in Pattern Recognition: 29th DAGM Symposium, Heidelberg, Germany, September 12-14, 2007. Proceedings 29. Springer, 2007, pp. 214–223

    2007

  18. [26]

    A spontaneous micro-expression database: Inducement, collection and baseline,

    Xiaobai Li, Tomas Pfister, Xiaohua Huang, Guoying Zhao, and Matti Pietik¨ainen, “A spontaneous micro-expression database: Inducement, collection and baseline,” in 2013 10th IEEE International Conference and Workshops on Automatic face and gesture recognition (fg) . IEEE, 2013, pp. 1–6

    2013

  19. [27]

    Casme ii: An improved spontaneous micro-expression database and the baseline evaluation,

    Wen-Jing Yan, Xiaobai Li, Su-Jing Wang, Guoying Zhao, Yong-Jin Liu, Yu-Hsin Chen, and Xiaolan Fu, “Casme ii: An improved spontaneous micro-expression database and the baseline evaluation,” PloS one, vol. 9, no. 1, pp. e86041, 2014

    2014

  20. [28]

    Samm: A spontaneous micro-facial movement dataset,

    Adrian K Davison, Cliff Lansley, Nicholas Costen, Kevin Tan, and Moi Hoon Yap, “Samm: A spontaneous micro-facial movement dataset,” IEEE transactions on affective computing , vol. 9, no. 1, pp. 116–129, 2016

    2016

  21. [29]

    Megc 2019–the second facial micro-expressions grand challenge,

    John See, Moi Hoon Yap, Jingting Li, Xiaopeng Hong, and Su- Jing Wang, “Megc 2019–the second facial micro-expressions grand challenge,” in 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2019) . IEEE, 2019, pp. 1–5

    2019

  22. [30]

    Dfme: A new benchmark for dynamic facial micro-expression recognition,

    Sirui Zhao, Huaying Tang, Xinglong Mao, Shifeng Liu, Yiming Zhang, Hao Wang, Tong Xu, and Enhong Chen, “Dfme: A new benchmark for dynamic facial micro-expression recognition,” IEEE Transactions on Affective Computing , 2023

    2023

  23. [31]

    Dy- namic micro-expression automatic recognition challenge on the fourth chinese conference on affective computing,

    Sirui Zhao, Huaying Tang, Xinglong Mao, and Shifeng Liu, “Dy- namic micro-expression automatic recognition challenge on the fourth chinese conference on affective computing,” https://mea-lab-421.github. io/CCAC-page/, July 10 2024

    2024

  24. [32]

    Diverse local facial behaviors learning from enhanced expression flow for microexpression recognition,

    Rongrong Ni, Biao Yang, Xu Zhou, Siyang Song, and Xiaofeng Liu, “Diverse local facial behaviors learning from enhanced expression flow for microexpression recognition,” Knowledge-Based Systems, vol. 275, pp. 110729, 2023

    2023

  25. [33]

    From macro to mi- cro: Boosting micro-expression recognition via pre-training on macro- expression videos,

    Hanting Li, Hongjing Niu, and Feng Zhao, “From macro to mi- cro: Boosting micro-expression recognition via pre-training on macro- expression videos,” arXiv preprint arXiv:2405.16451 , 2024

    2024 arXiv

  26. [34]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakr- ishna Vedantam, Devi Parikh, and Dhruv Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 618–626

    2017

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.