Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Online Micro-gesture Recognition Using Data Augmentation and Spatial-Temporal Attention

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This challenge report claims its data augmentation and spatial-temporal attention achieve an F1 of 38.03, a 37.9% gain, and first place in the IJCAI 2025 micro-gesture online recognition track.

desk verdict The paper's headline F1=38.03 belongs to the AdaTAD baseline, not the proposed method, which scores 33.44; the central claim collapses, though the ablation is honest. read the letter →

arxiv 2507.09512 v2 pith:ED6DGCEL submitted 2025-07-13 cs.CV

classification cs.CV
keywords micro-gestureonlinerecognitiondataaugmentationspatial-temporalattentiontemporalactiondetectionSMGdatasetIJCAIMiGAchallengeDyFADetVideoMAEv2
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 is the HFUT-VUT solution report for the IJCAI 2025 MiGA Micro-gesture Online Recognition track. It claims that a category-frequency-based annotation augmentation scheme and a spatial-temporal attention module on the detection head, built on the DyFADet architecture with VideoMAEv2-g features, improve classification and boundary localization, reaching an F1 of 38.03 and first place. The paper's own experiments tell a more measured story: the 38.03 score is listed under the AdaTAD baseline with VideoMAEv2-g, while the proposed pipeline scores 33.44, up from the DyFADet baseline's 27.78. The report's practical evidence is that both proposed modules reduce false negatives for short actions and lower confusion errors, per the included error analysis.

What carries the argument

The mechanism has two parts. The first is the category-frequency-based adaptive label augmentation: for each category $c$ with instance count $Z_c$ below a threshold $\alpha$, every annotation is repeated $R_c = \lfloor \log_2(\alpha/Z_c)\rfloor + 1$ times in the training set, rebalancing rare classes without duplicating video frames. The second is the spatial-temporal attention module inserted into the multi-scale TAD head of DyFADet, composed of a temporal attention branch using shared MLP on element-wise summed average- and max-pooled features, and a spatial attention branch applying a $7\times7$ convolution to concatenated average- and max-pooled channel features. These modules are meant to make the detector concentrate on informative frames and body regions, which the authors identify as the key to precise micro-gesture boundaries and category distinctions.

What would settle it

Rerun the submitted competition model and check the leaderboard entry: if the configuration that scored 38.03 did not include the proposed augmentation and attention modules, the abstract's attribution is false. Alternatively, run AdaTAD with VideoMAEv2-g features under the paper's protocol and verify whether it reproduces 38.03 without any of the proposed components.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that online micro-gesture recognition is improved by fixing two weaknesses: category imbalance in the training set and the detection head's weak spatial-temporal focus. The authors' augmentation counts per-category instances, sets a minimum threshold $\alpha$, and replicates rare-category annotations $R_c = \lfloor \log_2(\alpha/Z_c)\rfloor + 1$ times, avoiding pixel duplication. Their spatial-temporal attention module adds a temporal branch (shared MLP over average- and max-pooled features) and a spatial branch ($7\times7$ convolution over concatenated pooled features) to the multi-scale TAD head of DyFADet. The abstract's headline result, F1 38.03, is presented as the outcome of the full solution, although the ablation table records the proposed combination at 33.44 and the higher score under AdaTAD with the same backbone. The supporting evidence for the new modules is the improvement from 27.78 to 33.44 on the SMG dataset.

Load-bearing premise

The load-bearing premise is that the reported F1 of 38.03 was produced by the proposed data augmentation and spatial-temporal attention; the paper's own Table 2 assigns that score to the AdaTAD baseline with VideoMAEv2-g features, while the proposed pipeline scores 33.44, so if the table is correct the central claim collapses.

Editorial extensions

If this is right

  • If the ablation numbers hold, the augmentation alone improves the DyFADet baseline from 27.78 to 32.04 F1, so label replication is a low-cost way to counter rare-class imbalance in temporal detection.
  • The spatial-temporal attention alone raises the baseline to 32.69 F1, suggesting it contributes at least as much as the augmentation to the combined 33.44.
  • The combination of both modules outperforms either alone, indicating that rebalancing labels and refining features are complementary rather than redundant.
  • The reported test-set F1 of 38.03 would only be attributable to the proposed method if the winning submission used those modules; the paper's Table 2 instead lists that number for AdaTAD, so the gain the paper can claim is the 33.44 over 27.78.

Reading between the lines

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

  • A reader may infer that the first-place result reflects the VideoMAEv2-g features and AdaTAD detector rather than the two proposed components; resolving this requires the competition submission's configuration.
  • The annotation-level augmentation is a generic recipe that could be transferred to other label-imbalanced temporal action detection datasets and tested on standard benchmarks.
  • The spatial-temporal attention module is a generic plug-in; a systematic study varying only that module across multiple backbones and datasets would test its general value.
  • The error analysis points to confusion among visually similar gestures (e.g., rubbing hands vs. folding arms), suggesting that finger-level motion modeling is a likely next step.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents a solution for the IJCAI 2025 MiGA Challenge Micro-gesture Online Recognition track. It proposes an annotation-level data augmentation strategy that repeats rare-class instances according to a logarithmic factor, and a spatial-temporal attention module inserted into the multi-scale detection head of DyFADet. Experiments are conducted on the SMG dataset using VideoMAEv2-g features. The paper claims that the proposed method achieved an F1 score of 38.03, outperforming the previous state-of-the-art by 37.9% and ranking first in the competition. An ablation study shows improvements over a DyFADet baseline (27.78 to 33.44) when adding the proposed components, along with error analysis and sensitivity analysis. However, the paper's own Table 2 and Section 4.3 attribute the 38.03 score to an AdaTAD baseline without either proposed component, directly contradicting the central claim.

Significance. If the central claim were valid, the paper would be a modest but useful incremental contribution: a simple data augmentation scheme and a standard spatial-channel attention module applied to a recent temporal action detection framework, with a competition win as external validation. The paper does provide an ablation and an error analysis that suggest the proposed modules help relative to the DyFADet baseline. However, the load-bearing claim of state-of-the-art performance and first place is internally contradicted by the manuscript's own results. The proposed method's best reported F1 is 33.44, not 38.03, and the latter score is explicitly identified as a baseline that does not use the paper's contributions. The methodological novelty is also limited, as the spatial-temporal attention resembles conventional CBAM-style modules and the data augmentation is a simple log-based oversampling. The significance of the paper therefore hinges on a claim that the paper itself refutes.

major comments (3)
  1. [Abstract and Section 4.3, Table 2] The abstract and the contribution list state that 'Our solution achieved an F1 score of 38.03' and that 'our proposed method achieved an F1 score of 38.03', but Table 2 assigns the 38.03 F1 to AdaTAD [25] with VideoMAEv2-g and with neither proposed component (both entries marked as absent). Section 4.3 explicitly says: 'we adopted VideoMAEv2-g as the backbone and applied the method proposed by Liu et al. [25], achieving the best result with an F1 score of 38.03' and that this result is reported 'solely as a reference for future research.' The row labeled 'Ours' with both proposed components scores 33.44. Thus the paper's central empirical claim is internally contradicted: the competition-winning score is produced by a baseline that does not use the proposed data augmentation or spatial-temporal attention.
  2. [Section 4.3, Table 2] The claimed improvement of 37.9% over the previous state-of-the-art is not a controlled comparison and is not attributable to the proposed method. The 38.03 value is 37.9% higher than 27.57, which is the IJCAI 2024 rank-1 score of Wang et al. [23], but 38.03 is the AdaTAD baseline, not the proposed method. The proposed method's best score is 33.44, which is 21.3% higher than 27.57. Moreover, all results in Table 2 appear to be from single runs with no error bars, standard deviations, or significance tests, so even the ablation improvements (27.78 to 33.44) cannot be distinguished from training stochasticity.
  3. [Section 3.2 and Table 2] The data augmentation strategy in Eq. (1) depends on a free parameter, the minimum instance threshold alpha, but the paper provides no sensitivity analysis or guidance for setting alpha. Since the central claim of the paper relies on the effectiveness of this augmentation, the sensitivity of the result to alpha should be established. Additionally, the augmentation duplicates annotations of rare classes, which the authors acknowledge in Section 5 may cause overfitting; this risk is not empirically investigated. The ablation would be more convincing if it reported results across several values of alpha or across multiple random seeds.
minor comments (5)
  1. [Abstract and Conclusion] The abstract and the conclusion repeat the claim that 'our final model achieved a score of 38.03'; this needs to be corrected to reflect that 38.03 is attributed to the AdaTAD baseline in Section 4.3.
  2. [Section 3.1] The sentence 'each v_t in V is typically extracted using pre-trained video encodes' contains a typo: 'encodes' should be 'encoders'. Also, the notation 'N = {1, 2, ..., n}' conflates a set with an integer; this should be clarified.
  3. [Table 2] The use of the '%' symbol to indicate whether a component is used is unconventional and not explained in the caption; the caption or table header should define the notation clearly.
  4. [References] References [22] and [23] are incomplete; they lack full bibliographic details (e.g., page numbers or publisher information), which should be completed for archival quality.
  5. [Figure 2] The caption of Figure 2 is long and dense, and several module names in the figure (e.g., 'DynE Layer', 'Feature Fusion', 'Spatial-Temporal Attention') are not all explicitly defined in the main text before they appear; a more structured caption or a notation table would improve readability.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation; the 38.03 claim is an internal attribution inconsistency, not circular reasoning.

full rationale

The technical derivation is self-contained. Eq. (1) is a hand-crafted annotation-repetition schedule, and Eqs. (4)-(5) are standard CBAM-style attention definitions with pooling, MLP, sigmoid, and convolution; neither is fitted to nor derived from the target F1 score. The ablation in Table 2 provides independent content: adding both proposed components raises DyFADet from 27.78 to 33.44, so the proposed modules have a measurable effect that is not circularly defined. The serious problem is the abstract/conclusion claim that 'Our solution achieved an F1 score of 38.03.' Section 4.3 explicitly says the 38.03 best result was obtained by 'the method proposed by Liu et al. [25]' (AdaTAD) with no Data Augmentation and no Spatial-Temporal Attention, and that it is reported 'solely as a reference for future research.' In Table 2 the only 38.03 row is AdaTAD [25] with VideoMAEv2-g and neither proposed component, while 'Ours' with both components scores 33.44. This is an internal attribution/reporting inconsistency, not a circular reduction: no equation or fitted constant makes 38.03 equal to the proposed method's output by construction; the paper merely labels a baseline result as its own in Table 1. Self-citations such as Liu et al. [18] are contextual and not load-bearing. Therefore no significant circularity is present.

Assumptions & free parameters 1 free parameters · 2 assumptions · 0 invented entities

The central claim depends on the challenge leaderboard's evaluation and on the untested assumption that annotation-level duplication helps generalization. No free parameter is disclosed (alpha), and no invented entities are introduced.

free parameters (1)
  • minimum instance threshold alpha = not reported
    The augmentation rule in Eq. 1 uses alpha to decide which categories are rare and how many copies to add; no value is given.
assumptions (2)
  • domain assumption SMG test set and the Kaggle leaderboard provide a fair and complete evaluation of online micro-gesture recognition.
    The paper takes the leaderboard F1 as ground truth for its central claim without independent verification.
  • domain assumption Annotation-level replication of rare classes does not leak information or overfit the test set.
    The paper's augmentation is performed at annotation level; the paper itself notes in the conclusion that overfitting may result.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Online Micro-gesture Recognition Using Data Augmentation and Spatial-Temporal Attention." pith.science (2026). https://pith.science/paper/ED6DGCEL

@misc{pith2026250709512,
  author       = {Pith},
  title        = {Pith review of: Online Micro-gesture Recognition Using Data Augmentation and Spatial-Temporal Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ED6DGCEL}},
  note         = {Machine review of arXiv:2507.09512}
}
read the original abstract

In this paper, we introduce the latest solution developed by our team, HFUT-VUT, for the Micro-gesture Online Recognition track of the IJCAI 2025 MiGA Challenge. The Micro-gesture Online Recognition task is a highly challenging problem that aims to locate the temporal positions and recognize the categories of multiple micro-gesture instances in untrimmed videos. Compared to traditional temporal action detection, this task places greater emphasis on distinguishing between micro-gesture categories and precisely identifying the start and end times of each instance. Moreover, micro-gestures are typically spontaneous human actions, with greater differences than those found in other human actions. To address these challenges, we propose hand-crafted data augmentation and spatial-temporal attention to enhance the model's ability to classify and localize micro-gestures more accurately. Our solution achieved an F1 score of 38.03, outperforming the previous state-of-the-art by 37.9%. As a result, our method ranked first in the Micro-gesture Online Recognition track.

Figures

Figures reproduced from arXiv: 2507.09512 by the authors.

Figure 1
Figure 1. Micro-Gesture Category Distribution on the training set of the SMG dataset. The blue side is the SMG dataset, while the orange side is the data after applying data augmentation. The number of samples in underrepresented categories has significantly increased following augmentation. 2.3. Temporal Action Detection Temporal Action Detection (TAD) [24, 25, 26, 27] aims to locate and classify all actions in untrimmed vid… view at source ↗
Figure 2
Figure 2. Overview of the proposed method. The Feature Encoder consists of downsampling and DynE layers. The multi-scale fusion module in the Multi-scale Spatial-Temporal Attention TAD Head comprises Feature Fusion and Spatial-Temporal Attention. First, we count the number of instances for each gesture category in the training set and define a minimum instance threshold 𝛼. For categories with fewer than 𝛼 instances, we calcul… view at source ↗
Figure 3
Figure 3. Error Analysis of the baseline and the proposed method using a standard TAD evaluation toolkit [34]. Figures (a), (c), and (e) show the performance metrics of the baseline, while Figures (b), (d), and (f) illustrate the corresponding results achieved by our method. Ground Truth: Rubbing hands and crossing finger Prediction: Folding arms Ground Truth: Scratching or touching facial parts other than eyes Prediction: Pl… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Typical examples of classification errors. Left figure: The model incorrectly predicted “Rubbing hands and crossing fingers” as “Folding arms”; Right figure: The model incorrectly predicted “Scratching or touching facial parts other than eyes” as “Playing or adjusting …

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MAC 2026: Advancing Micro-Action Analysis Towards Fine-Grained Understanding

    cs.CV 2026-07 conditional novelty 3.0 of 10

    MAC 2026 reports a three-track micro-action challenge, adding a fine-grained MLLM-based understanding track evaluated on MA-Bench, with top-3 leaderboard results for each track.

Reference graph

Works this paper leans on

34 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [25]

    Liu, C.-L

    S. Liu, C.-L. Zhang, C. Zhao, B. Ghanem, End-to-end temporal action detection with 1b parameters across 1000 frames, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18591–18601

  2. [23]

    Y. Wang, L. Kerui, H. Huang, Z. Xia, Micro-gesture online recognition with dual-stream multi-scale transformer in long videos, MiGA@ IJCAI (2024)

  3. [1]

    X. Liu, H. Shi, H. Chen, Z. Yu, X. Li, G. Zhao, imigue: An identity-free video dataset for micro- gesture understanding and emotion analysis, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 10631–10642

  4. [2]

    H. Chen, H. Shi, X. Liu, X. Li, G. Zhao, Smg: A micro-gesture dataset towards spontaneous body gestures for emotional stress state analysis, International Journal of Computer Vision 131 (2023) 1346–1366

  5. [3]

    H. Chen, X. Liu, X. Li, H. Shi, G. Zhao, Analyze spontaneous gestures for emotional stress state recognition: A micro-gesture dataset and analysis with deep learning, in: 2019 14th IEEE International Conference on Automatic Face & Gesture Recognition, 2019, pp. 1–8

  6. [4]

    P. Liu, G. Dong, D. Guo, K. Li, F. Li, X. Yang, M. Wang, X. Ying, A survey on fmri-based brain decoding for reconstructing multimodal stimuli, arXiv preprint arXiv:2503.15978 (2025)

  7. [5]

    Z. Wei, J. Zhang, Z. Lin, J.-Y. Lee, N. Balasubramanian, M. Hoai, D. Samaras, Learning visual emotion representations from web data, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020, pp. 13106–13115

  8. [6]

    R. Gao, X. Liu, B. Xing, Z. Yu, B. W. Schuller, H. Kälviäinen, Identity-free artificial emotional intelligence via micro-gesture understanding, arXiv preprint arXiv:2405.13206 (2024)

Show all 34 references
  1. [7]

    D. Guo, K. Li, B. Hu, Y. Zhang, M. Wang, Benchmarking micro-action recognition: Dataset, methods, and applications, IEEE Transactions on Circuits and Systems for Video Technology 34 (2024) 6238–6252

  2. [8]

    D. Guo, X. Li, K. Li, H. Chen, J. Hu, G. Zhao, Y. Yang, M. Wang, Mac 2024: Micro-action analysis grand challenge, in: Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp. 11304–11305

  3. [9]

    J. Zhao, F. Wang, K. Li, Y. Wei, S. Tang, S. Zhao, X. Sun, Temporal-frequency state space duality: An efficient paradigm for speech emotion recognition, in: ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2025, pp. 1–5

  4. [10]

    K. Li, D. Guo, G. Chen, C. Fan, J. Xu, Z. Wu, H. Fan, M. Wang, Prototypical calibrating ambiguous samples for micro-action recognition, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 2025, pp. 4815–4823

  5. [11]

    K. Li, D. Guo, G. Chen, F. Liu, M. Wang, Data augmentation for human behavior analysis in multi- person conversations, in: Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 9516–9520

  6. [12]

    F. Wang, D. Guo, K. Li, Z. Zhong, M. Wang, Frequency decoupling for motion magnification via multi-level isomorphic architecture, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 18984–18994

  7. [13]

    F. Wang, D. Guo, K. Li, M. Wang, Eulermormer: Robust eulerian motion magnification via dynamic filtering within transformer, in: Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 2024, pp. 5345–5353

  8. [14]

    K. Li, X. Peng, D. Guo, X. Yang, M. Wang, Repetitive action counting with hybrid temporal relation modeling, IEEE Transactions on Multimedia (2025)

  9. [15]

    F. Wang, K. Li, Y. Nie, Z. Duan, P. Zou, Z. Wu, Y. Wang, Y. Wei, Exploiting ensemble learning for cross-view isolated sign language recognition, arXiv preprint arXiv:2502.02196 (2025)

  10. [16]

    K. Li, D. Guo, G. Chen, X. Peng, M. Wang, Joint skeletal and semantic embedding loss for micro-gesture classification, arXiv preprint arXiv:2307.10624 (2023)

  11. [17]

    G. Chen, F. Wang, K. Li, Z. Wu, H. Fan, Y. Yang, M. Wang, D. Guo, Prototype learning for micro-gesture classification, arXiv preprint arXiv:2408.03097 (2024)

  12. [18]

    P. Liu, F. Wang, K. Li, G. Chen, Y. Wei, S. Tang, Z. Wu, D. Guo, Micro-gesture online recognition using learnable query points, arXiv preprint arXiv:2407.04490 (2024)

  13. [19]

    J. Gu, K. Li, F. Wang, Y. Wei, Z. Wu, H. Fan, M. Wang, Motion matters: Motion-guided modu- lation network for skeleton-based micro-action recognition, in: Proceedings of the 33rd ACM International Conference on Multimedia, 2025

  14. [20]

    S. Sun, D. Liu, J. Dong, X. Qu, J. Gao, X. Yang, X. Wang, M. Wang, Unified multi-modal unsupervised representation learning for skeleton-based action understanding, in: Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 2973–2984

  15. [21]

    K. Li, P. Liu, D. Guo, F. Wang, Z. Wu, H. Fan, M. Wang, Mmad: Multi-label micro-action detection in videos, arXiv preprint arXiv:2407.05311 (2024)

  16. [22]

    X. Guo, W. Peng, H. Huang, Z. Xia, Micro-gesture online recognition with graph-convolution and multiscale transformers for long sequence., in: MiGA@ IJCAI, 2023

  17. [24]

    J. Tan, X. Zhao, X. Shi, B. Kang, L. Wang, Pointtad: Multi-label temporal action detection with learnable query points, Advances in Neural Information Processing Systems 35 (2022) 15268–15280

  18. [26]

    Zhang, J

    C.-L. Zhang, J. Wu, Y. Li, Actionformer: Localizing moments of actions with transformers, in: Proceedings of the European Conference on Computer Vision, Springer, 2022, pp. 492–510

  19. [27]

    L. Yang, Z. Zheng, Y. Han, H. Cheng, S. Song, G. Huang, F. Li, Dyfadet: Dynamic feature aggregation for temporal action detection, in: Proceedings of the European Conference on Computer Vision, Springer, 2024, pp. 305–322

  20. [28]

    Z. Tong, Y. Song, J. Wang, L. Wang, Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training, Advances in Neural Information Processing Systems 35 (2022) 10078–10093

  21. [29]

    L. Wang, B. Huang, Z. Zhao, Z. Tong, Y. He, Y. Wang, Y. Wang, Y. Qiao, Videomae v2: Scaling video masked autoencoders with dual masking, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 14549–14560

  22. [30]

    Tirupattur, K

    P. Tirupattur, K. Duarte, Y. S. Rawat, M. Shah, Modeling multi-label action dependencies for temporal action localization, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 1460–1470

  23. [31]

    R. Dai, S. Das, K. Kahatapitiya, M. S. Ryoo, F. Brémond, Ms-tct: Multi-scale temporal convtrans- former for action detection, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 20041–20051

  24. [32]

    D. Shi, Y. Zhong, Q. Cao, L. Ma, J. Li, D. Tao, Tridet: Temporal action detection with relative boundary modeling, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18857–18866

  25. [33]

    Carreira, A

    J. Carreira, A. Zisserman, Quo vadis, action recognition? a new model and the kinetics dataset, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 6299–6308

  26. [34]

    Alwassel, F

    H. Alwassel, F. C. Heilbron, V. Escorcia, B. Ghanem, Diagnosing error in temporal action detectors, in: Proceedings of the European Conference on Computer Vision, 2018, pp. 256–272

Pith tools

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