Pith. sign in

REVIEW 3 major objections 4 minor 31 references

A lightweight temporal convolutional network matches a graph-based skeleton model for body-emotion recognition while using 79% fewer parameters and running 12.5× faster.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 09:16 UTC pith:JKKNSJIS

load-bearing objection A solid but under-supported efficiency claim: the TCN-vs-graph benchmark is useful, but the 'practical alternative' conclusion needs fold-level data and a pre-specified margin. the 3 major comments →

arxiv 2607.20820 v2 pith:JKKNSJIS submitted 2026-07-23 cs.AI

Efficient and Interpretable Body-Based Emotion Recognition with Lightweight Temporal Convolutional Networks

Christian Arzate Cruz , Stefanos Gkikas , Houshyar Asadi This is my paper
classification cs.AI
keywords body-based emotion recognitiontemporal convolutional networksskeleton-based recognitionDIEM-Amodel interpretabilitycomputational efficiencysaliency analysisbody-region analysis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks whether a small temporal convolutional network can replace a heavier graph-based skeleton model for classifying emotions from body motion, without sacrificing accuracy. It compares several TCN sizes against a graph baseline (G-TSG) on the DIEM-A dataset. It finds that the mid-size TCN-Base stays within 1.58 accuracy points and 1.25 macro-F1 points of the graph model, while using 79.18% fewer parameters and reducing inference latency by roughly 12.5×. The authors also probe which body regions carry emotion information, finding that upper-body motion is the strongest standalone cue, but that regions contribute differently across emotions. The paper concludes that lightweight TCNs are a practical, efficient alternative for real-time affective systems and that combining multiple interpretability methods gives a clearer picture than any single one.

Core claim

On the DIEM-A body-emotion dataset, a deliberately simple temporal-only convolutional network (TCN-Base, 286,648 parameters) achieves 25.53% accuracy and 23.96% macro-F1, versus 27.11% and 25.21% for the graph-based G-TSG baseline (1,376,751 parameters). The accuracy gap is only 1.58 points and the macro-F1 gap 1.25 points, and paired fold-level tests (Wilcoxon W=17, p=0.322; t(9)=1.196, p=0.262 for accuracy) do not detect a statistically significant difference. Meanwhile TCN-Base uses 79.18% fewer parameters, 98.03% fewer MACs, and roughly 12.5× lower per-clip latency (1.107 ms vs 13.851 ms). The paper also reports that a TCN trained only on upper-body input retains most of the full-body pe

What carries the argument

The central object is the dilated residual temporal convolutional block: a flattened skeleton-joint feature sequence (C×V channels over T frames) is processed by stacked 1D convolutions with exponentially increasing dilations, batch normalization, ReLU, dropout, and residual connections. This removes explicit spatial graph operations, allowing a tiny model to capture multi-scale temporal motion patterns in parallel. The comparison rests on paired leave-performer-out folds, with accuracy and macro-F1 evaluated fold-wise and tested with Wilcoxon and paired-t tests, while latency is measured with CUDA-synchronized batch-1 forward passes.

Load-bearing premise

The paper interprets the statistically non-significant fold-level accuracy gap (1.58 points) as evidence that TCN-Base is a practical low-latency alternative, even though 10 folds provide limited power to detect a true difference of a few points at this task's ~25% accuracy level.

What would settle it

Run the same comparison with more leave-performer-out folds (e.g., increasing effective test set size) or with a held-out performer cohort, and measure whether the accuracy gap exceeds ~2-3 points and becomes statistically significant; if a real gap of several points appears at this low accuracy, the practical-equivalence claim weakens substantially.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Real-time affective systems that previously relied on graph-based skeleton models can switch to TCN-Base and cut classifier latency from ~13.9 ms to ~1.1 ms per clip with negligible mean-performance loss.
  • A 286K-parameter temporal model is sufficient to capture usable emotion cues from 6D joint rotations, suggesting that explicit body-graph structure is not necessary for this task's performance level.
  • The region-specific results imply that future body-emotion models can prioritize upper-body joints (head, torso, arms/hands) for compact sensors or partial-occlusion settings.
  • The disagreement between region-only training (torso weak) and gradient saliency (torso strongest) implies that interpretability claims must specify which question they answer: standalone sufficiency, perturbation dependence, or local sensitivity.
  • Per-emotion regional patterns (arms/hands for anger and fear, legs for shame, torso for sadness and gratitude) suggest that a single global region ranking is insufficient for designing emotion-aware interaction systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the 1.58-point gap is real but small, a larger sample of performers or folds might reveal statistical significance; the paper's own caveat that it does 'not establish statistical equivalence' already marks this interpretive boundary.
  • The 12.5× latency figure excludes upstream pose estimation; in a full pipeline the end-to-end speedup will be smaller, and the practical gain depends on the ratio of pose-estimation to classifier cost, which the paper does not measure.
  • The same TCN architecture could be tested on other skeleton-emotion datasets (e.g., cross-cultural or clinical populations) to check whether the upper-body dominance and torso-saliency dissociation generalize beyond DIEM-A.
  • A parameter-matched graph model (same size as TCN-Base) would isolate whether the performance gap stems from architecture or capacity, a comparison the paper explicitly leaves to future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This paper evaluates a family of lightweight temporal convolutional networks (TCNs) against a graph-based G-TSG baseline on the DIEM-A body-emotion benchmark. It reports that TCN-Base remains within 1.58 accuracy points and 1.25 macro-F1 points of G-TSG while using 79.18% fewer parameters and approximately 12.5× lower latency, and that the performance difference is not statistically significant under paired fold-level tests. The paper then analyzes body-region evidence using region-specific TCN training, zero-based occlusion, and G-TSG gradient saliency, concluding that upper-body motion is the strongest standalone regional cue, that regional usefulness varies across emotions, and that the three interpretability methods capture distinct aspects of model behavior.

Significance. If the efficiency result holds, the paper offers a practically relevant data point: a 286K-parameter TCN approaches a 1.38M-parameter graph model on a 12-way body-emotion task with roughly two orders of magnitude fewer MACs and an order-of-magnitude latency reduction. The latency measurement is carefully described, the authors explicitly avoid overclaiming statistical equivalence, and the interpretability analysis is more disciplined than typical saliency-only studies: it separates sufficiency, perturbation dependence, and local sensitivity, and honestly treats zero-based occlusion as diagnostic. The main weakness is that the central 'practical low-latency alternative' claim rests on a non-significant difference from an underpowered 10-fold paired test, without confidence intervals, a pre-specified non-inferiority margin, or seed-level variability. The absence of code and seeds further limits reproducibility and verification of the headline gap.

major comments (3)
  1. [§IV.A, Table II] The claim that TCN-Base is a 'practical low-latency alternative' is load-bearing but not yet established. The paired tests have only n=10 folds. From the reported paired t-statistic t(9)=1.196 and a mean accuracy difference of 1.58 points, the fold-difference SD is approximately 4.17 points, yielding a 95% CI for the true mean difference of roughly (−1.4, +4.6) accuracy points. At ~25.5% mean accuracy, a 4.6-point gap is a material relative shortfall for a 12-way task. The Wilcoxon test (W=17, p=0.322) is similarly insensitive. The authors correctly disclaim statistical equivalence, but the 'practical alternative' wording leans on the point estimate. Please report fold-level paired differences, the 95% CI for the mean difference, and either a pre-specified non-inferiority margin or an explicit statement of the largest accuracy gap that would still be practically acceptable.
  2. [§III.D, §IV.A] No code or seeds are provided, and the full-body TCN-Base accuracy differs across tables: 25.53% in Table II, 24.90% in Table III, and 27.08% in Table IV. These are described as separate controlled runs, but without seed reporting the reader cannot tell whether the headline 1.58-point gap is a favorable realization. Please report the random seed(s), per-fold results, and repeated-seed summary statistics for the primary G-TSG versus TCN-Base comparison, and for the region-specific TCN runs. This is essential for assessing the stability of the efficiency claim and the regional hierarchy.
  3. [§IV.B, Table III] The regional analysis is largely sound, but the statistical support is thinner than the text suggests. Only macro-F1 is tested with Holm-adjusted Wilcoxon tests; accuracy differences are not tested. Given that the full-body reference itself varies between tables (24.90 vs. 25.53), the reported drops could change with seeds. I recommend reporting fold-level paired tests for both accuracy and macro-F1, or at least stating clearly that accuracy differences were not statistically tested. This is not fatal to the regional conclusions, but it would make the 'strongest standalone cue' claim more robust.
minor comments (4)
  1. [Abstract, §IV.A] The abstract and Section IV.A give slightly different latency-reduction numbers ('approximately 12.5×' vs. 'approximately 12×'). Table II implies 13.851/1.107 = 12.51×; please unify the wording.
  2. [§III.B] Eq. (3) writes X ∈ R^{C×T×V}, while the TCN input is described as being reshaped to (CV)×T. Making this reshape explicit as an equation would improve clarity.
  3. [§IV.C] The text says 'a full confusion-matrix analysis is left to supplementary material,' but no supplementary file is mentioned in the submission. Please either include the supplementary material or remove the reference.
  4. [Table V] Table V compares torso, arms/hands, and legs but omits head, even though Figure 2 reports head saliency. If head is excluded for space, say so; otherwise include a head row for completeness.

Circularity Check

0 steps flagged

No significant circularity: empirical benchmark, self-citations are contextual only.

full rationale

The paper's central claims are direct empirical measurements: TCN-Base is within 1.58 accuracy points of G-TSG, uses 79.18% fewer parameters, and is about 12.5x faster (Table II). These numbers are read off benchmark runs, not derived from any fitted parameter or defining equation. No target quantity is defined in terms of another result: the paired tests compare measured fold-wise differences, and the paper explicitly disclaims statistical equivalence or non-inferiority. The body-region analyses (region-specific training, occlusion, saliency) are post hoc observations about trained models, presented with explicit caveats about what each method can and cannot support; they are not predictions derived from fitted inputs. The authors' self-citations [4]-[10] appear in the introduction and related work as contextual references to prior affective-computing work and are not load-bearing for any reported result or architecture choice. The discussion of underpowered statistical tests and unstable full-body reference values across tables is a correctness/interpretation concern, not circularity. No uniqueness theorem is imported, no ansatz is smuggled in via citation, and no known result is merely renamed. The paper is self-contained as an empirical comparison, so the circularity score is low.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

All central results are empirical and depend on hand-chosen architecture and analysis choices; no new entities are introduced. The main untested premises are the interpretability proxies (gradient saliency and region-sufficiency training) and the integrity of the external benchmark pipeline.

free parameters (5)
  • TCN family hyperparameters (hidden width d, dilation schedule, kernel size 3, dropout 0.5) = d in {64,96,128,192,128}; dilations in {[1,2,4], [1,2,4,8], [1,2,4,8,16]}
    Chosen by hand in Table I; the efficiency/performance trade-off and the selection of TCN-Base for all follow-up analyses depend on these values.
  • Clip length T = 64 frames
    Fixed input length from the official preprocessing pipeline; every model sees the same temporal window.
  • Body-region partition = head=[6,7,8]; torso=[0..5]; arms/hands=[9..16]; legs=[17..24]; upper body=[0..16]; lower body=[0,1,17..24]
    Hand-defined joint groupings in Section III-Cd; all region-sufficiency and saliency conclusions are conditional on this particular partition.
  • Saliency aggregation and normalization = abs-gradient summed over channels/time, averaged over joints per region, normalized to sum 100% per sample
    The G-TSG saliency ranking (torso > legs > head > arms/hands) is a consequence of these aggregation choices, which are not uniquely forced by the data.
  • Statistical thresholds and corrections = alpha=0.05, Holm correction
    Holm-adjusted paired Wilcoxon tests in Table III determine which region models are declared significantly worse; conclusions about regional hierarchy depend on this choice.
axioms (5)
  • domain assumption DIEM-A labels and 6D joint-rotation preprocessing are correct and identical across models and folds
    Section III-B states that all models share the same preprocessed skeleton sequences and fold definitions; any label or preprocessing error would affect every comparison.
  • domain assumption Leave-performer-out folds prevent performer leakage
    Section III-D relies on fold construction from the official code; if performers leak across train/test, the reported accuracy and F1 would be inflated.
  • domain assumption Input-gradient saliency is a meaningful measure of local sensitivity
    Section III-E aggregates absolute input gradients to rank body regions; this is a standard but unvalidated proxy, and the paper itself notes that different methods answer different questions.
  • domain assumption Region-only TCN accuracy measures standalone discriminative sufficiency
    Section II-D/III-E equates region-only performance with the information available in that region; an isolated region could underperform because the classifier lacks coordination context, not because the region carries no signal.
  • domain assumption The official DIEM-A baseline configuration is a sound, complete training protocol
    Section III-B says optimizer settings, training schedule, and evaluation follow the DIEM-A baseline configuration without listing them; reproducibility depends on that external, uncited code.

pith-pipeline@v1.3.0-alltime-deepseek · 11372 in / 15287 out tokens · 157907 ms · 2026-08-01T09:16:03.392406+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Efficient and Interpretable Body-Based Emotion Recognition with Lightweight Temporal Convolutional Networks." pith.science (2026). https://pith.science/paper/JKKNSJIS

@misc{pith2026260720820,
  author       = {Pith},
  title        = {Pith review of: Efficient and Interpretable Body-Based Emotion Recognition with Lightweight Temporal Convolutional Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JKKNSJIS}},
  note         = {Machine review of arXiv:2607.20820}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Body-based emotion recognition is important for real-time affective systems, but graph-based skeleton models can be computationally expensive. This paper studies whether lightweight temporal convolutional networks (TCNs) can provide an efficient and interpretable alternative for body-based emotion classification. We evaluate a family of TCN models on DIEM-A and compare them with a graph-based time-series graph (G-TSG) baseline using accuracy, macro-F1, parameter count, and inference latency. Although G-TSG achieves the highest mean performance, TCN-Base remains within $1.58$ accuracy points and $1.25$ macro-F1 points while using $79.18\%$ fewer parameters and reducing classifier latency by approximately $12.5\times$. We also analyze body-region contributions using region-specific TCN models, zero-based occlusion, and G-TSG gradient saliency. The results show that upper-body motion provides the strongest standalone regional cue, that the usefulness of body regions varies across emotions, and that different interpretability methods capture distinct aspects of model behavior. These findings suggest that lightweight TCNs can support efficient body-based emotion recognition while also providing practical insight into how motion cues contribute to classification.

Figures

Figures reproduced from arXiv: 2607.20820 by Christian Arzate Cruz, Houshyar Asadi, Stefanos Gkikas.

Figure 1
Figure 1. Figure 1: Per-emotion recall for TCN-Base models trained with different body-region inputs. Emotion-relevant information is not [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: G-TSG input-gradient saliency by body region and ground-truth emotion. Saliency is consistently highest for the torso [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 2 linked inside Pith

  1. [1]

    Survey on emotional body gesture recognition,

    F. Noroozi, C. A. Corneanu, D. Kami ´nska, T. Sapi ´nski, S. Escalera, and G. Anbarjafari, “Survey on emotional body gesture recognition,”IEEE transactions on affective computing, vol. 12, no. 2, pp. 505–523, 2018

    2018

  2. [2]

    J. F. Iaccino,Left brain-right brain differences: Inquiries, evidence, and new approaches. Psychology Press, 2014

    2014

  3. [3]

    The body in language,

    H. Ruthrof, “The body in language,” 2015

    2015

  4. [4]

    A Unified Tokenization Framework for Pain Recognition using Heterogeneous 3D Modalities,

    S. Gkikas, C. A. Cruz, V . Becchetti, M. U. Khan, A. Giuseppi, and R. F. Rojas, “A Unified Tokenization Framework for Pain Recognition using Heterogeneous 3D Modalities,” inProceedings of the 28th ACM International Conference on Multimodal Interaction. Association for Computing Machinery, 2026

    2026

  5. [5]

    ReFace: Reorganizing Facial Spatiotemporal Representations for Improved Pain Assessment,

    S. Gkikas, Y . Fang, C. A. Cruz, M. U. Khan, and R. F. Rojas, “ReFace: Reorganizing Facial Spatiotemporal Representations for Improved Pain Assessment,” in2026 14th International Conference on Affective Com- puting and Intelligent Interaction (ACII). IEEE, 2026

    2026

  6. [6]

    Towards a Unified Modality-Agnostic Multimodal Framework for Cognitive Workload Assessment,

    S. Gkikas, C. A. Cruz, C. Joseph, G. Giannakakis, and R. F. Rojas, “Towards a Unified Modality-Agnostic Multimodal Framework for Cognitive Workload Assessment,” in2026 14th International Conference on Affective Computing and Intelligent Interaction (ACII). IEEE, 2026

    2026

  7. [7]

    Empathetic robots using empathy classifiers in hri settings,

    C. Arzate Cruz, E. C. Montiel-V ´azquez, C. Maeda, D. Lam, and R. Gomez, “Empathetic robots using empathy classifiers in hri settings,” in2025 20th ACM/IEEE International Conference on Human-Robot Interaction (HRI), 2025, pp. 1211–1215

    2025

  8. [8]

    When and how to express empathy in human-robot interaction scenarios,

    C. Arzate Cruz, E. C. Montiel-Vazquez, C. Maeda, and R. Gomez, “When and how to express empathy in human-robot interaction scenarios,” in2025 34th IEEE International Conference on Robot and Human Interactive Communication (RO-MAN). IEEE, 2025, pp. 1070–1077

    2025

  9. [9]

    Efficient emotion-aware iconic gesture prediction for robot co-speech,

    E. C. Montiel-Vazquez, C. A. Cruz, S. Gkikas, T. Kassiotis, G. Gi- annakakis, and R. Gomez, “Efficient emotion-aware iconic gesture prediction for robot co-speech,”arXiv preprint arXiv:2604.11417, 2026

    Pith/arXiv arXiv 2026

  10. [10]

    Data aug- mentation for 3dmm-based arousal-valence prediction for hri,

    C. Arzate Cruz, Y . Sechayk, T. Igarashi, and R. Gomez, “Data aug- mentation for 3dmm-based arousal-valence prediction for hri,” in2024 33rd IEEE International Conference on Robot and Human Interactive Communication (ROMAN). IEEE, 2024, pp. 2015–2022

    2024

  11. [11]

    Skeleton-based emotion recognition based on two-stream self-attention enhanced spatial-temporal graph convolutional network,

    J. Shi, C. Liu, C. T. Ishi, and H. Ishiguro, “Skeleton-based emotion recognition based on two-stream self-attention enhanced spatial-temporal graph convolutional network,”Sensors, vol. 21, no. 1, p. 205, 2020

    2020

  12. [12]

    Skeleton-based explainable bodily expressed emotion recognition through graph con- volutional networks,

    E. Ghaleb, A. Mertens, S. Asteriadis, and G. Weiss, “Skeleton-based explainable bodily expressed emotion recognition through graph con- volutional networks,” in2021 16th IEEE International Conference on Automatic Face and Gesture Recognition (FG 2021). IEEE, 2021, pp. 1–8

    2021

  13. [13]

    Looking into gait for perceiving emotions via bilateral posture and movement graph convolutional networks,

    Y . Zhai, G. Jia, Y .-K. Lai, J. Zhang, J. Yang, and D. Tao, “Looking into gait for perceiving emotions via bilateral posture and movement graph convolutional networks,”IEEE Transactions on Affective Computing, vol. 15, no. 3, pp. 1634–1648, 2024

    2024

  14. [14]

    Multiscale 3d-shift graph convolution network for emotion recognition from human actions,

    H. Shi, W. Peng, H. Chen, X. Liu, and G. Zhao, “Multiscale 3d-shift graph convolution network for emotion recognition from human actions,” IEEE Intelligent Systems, vol. 37, no. 4, pp. 103–110, 2022

    2022

  15. [15]

    Asian emotional body movement database: Diverse intercultural e-motion database of asian performers (DIEM-A),

    M. Cheng, C.-h. Tseng, K. Fujiwara, V . Schneider, and Y . Kitamura, “Asian emotional body movement database: Diverse intercultural e-motion database of asian performers (DIEM-A),” in2025 13th International Conference on Affective Computing and Intelligent Interaction (ACII). IEEE, 2025

    2025

  16. [16]

    Affect recognition from face and body: early fusion vs. late fusion,

    H. Gunes and M. Piccardi, “Affect recognition from face and body: early fusion vs. late fusion,” in2005 IEEE international conference on systems, man and cybernetics, vol. 4. IEEE, 2005, pp. 3437–3443

    2005

  17. [17]

    Technique for automatic emotion recognition by body gesture analysis,

    D. Glowinski, A. Camurri, G. V olpe, N. Dael, and K. Scherer, “Technique for automatic emotion recognition by body gesture analysis,” in2008 IEEE Computer society conference on computer vision and pattern recognition workshops. IEEE, 2008, pp. 1–6

    2008

  18. [18]

    Toward a minimal representation of affective gestures,

    D. Glowinski, N. Dael, A. Camurri, G. V olpe, M. Mortillaro, and K. Scherer, “Toward a minimal representation of affective gestures,” IEEE Transactions on Affective Computing, vol. 2, no. 2, pp. 106–118, 2011

    2011

  19. [19]

    Towards a minimal representation of affective gestures,

    D. Glowinski, M. Mortillaro, K. Scherer, N. Dael, G. V olpe, and A. Camurri, “Towards a minimal representation of affective gestures,” in 2015 International Conference on Affective Computing and Intelligent Interaction (ACII). IEEE, 2015, pp. 498–504

    2015

  20. [20]

    A study on emotion recognition from body gestures using kinect sensor,

    S. Saha, S. Datta, A. Konar, and R. Janarthanan, “A study on emotion recognition from body gestures using kinect sensor,” in2014 international conference on communication and signal processing. IEEE, 2014, pp. 056–060

    2014

  21. [21]

    Multimodal emotion recognition in speech-based interaction using facial expression, body gesture and acoustic analysis,

    L. Kessous, G. Castellano, and G. Caridakis, “Multimodal emotion recognition in speech-based interaction using facial expression, body gesture and acoustic analysis,”Journal on Multimodal User Interfaces, vol. 3, no. 1, pp. 33–48, 2010

    2010

  22. [22]

    Skeleton based action recognition with convolutional neural network,

    Y . Du, Y . Fu, and L. Wang, “Skeleton based action recognition with convolutional neural network,” in2015 3rd IAPR Asian conference on pattern recognition (ACPR). IEEE, 2015, pp. 579–583

    2015

  23. [23]

    Spatial temporal graph convolutional networks for skeleton-based action recognition,

    S. Yan, Y . Xiong, and D. Lin, “Spatial temporal graph convolutional networks for skeleton-based action recognition,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, no. 1, 2018

    2018

  24. [24]

    Skeleton-based action recognition with directed graph neural networks,

    L. Shi, Y . Zhang, J. Cheng, and H. Lu, “Skeleton-based action recognition with directed graph neural networks,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 7912– 7921

    2019

  25. [25]

    Temporal convolutional networks for action segmentation and detection,

    C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager, “Temporal convolutional networks for action segmentation and detection,” in proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 156–165

    2017

  26. [26]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,”arXiv preprint arXiv:1803.01271, 2018

    Pith/arXiv arXiv 2018

  27. [27]

    Comparison between recurrent networks and temporal convolutional networks approaches for skeleton-based action recognition,

    M. Nan, M. Tr˘asc˘au, A. M. Florea, and C. C. Iacob, “Comparison between recurrent networks and temporal convolutional networks approaches for skeleton-based action recognition,”Sensors, vol. 21, no. 6, p. 2051, 2021

    2051

  28. [28]

    The body action and posture coding system (bap): Development and reliability,

    N. Dael, M. Mortillaro, and K. R. Scherer, “The body action and posture coding system (bap): Development and reliability,”Journal of Nonverbal Behavior, vol. 36, no. 2, pp. 97–121, 2012

    2012

  29. [29]

    Emotion expression in body action and posture

    ——, “Emotion expression in body action and posture.”Emotion, vol. 12, no. 5, p. 1085, 2012

    2012

  30. [30]

    Cross-cultural differences in recognizing affect from body posture,

    A. Kleinsmith, P. R. De Silva, and N. Bianchi-Berthouze, “Cross-cultural differences in recognizing affect from body posture,”Interacting with computers, vol. 18, no. 6, pp. 1371–1389, 2006

    2006

  31. [31]

    Affective body expression perception and recognition: A survey,

    A. Kleinsmith and N. Bianchi-Berthouze, “Affective body expression perception and recognition: A survey,”IEEE Transactions on Affective Computing, vol. 4, no. 1, pp. 15–33, 2012

    2012