Pith. sign in

REVIEW 2 major objections 4 minor 58 references

SHARDeg: A Benchmark for Skeletal Human Action Recognition in Degraded Scenarios

T0 review · 2 major / 4 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 At 3 FPS, a rough-path model overtakes the state of the art in skeletal action recognition.

desk verdict Useful benchmark, but the LogSigRNN re-ranking needs an input-matched test before it can support the architecture claim. read the letter →

arxiv 2505.18048 v2 pith:D7ACY7HJ submitted 2025-05-23 cs.CV

Simon Malzard , Nitish Mital , Richard Walters , Victoria Nockles , Raghuveer Rao , Celso M. De Melo This is my paper · ORCID
classification cs.CV
keywords skeletalhumanactionrecognitionrobustnessbenchmarkdatadegradationroughpaththeorylog-signaturetransformNTU-RGB+D-120framerateinterpolationmitigation
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 introduces the first systematic benchmark for skeletal human action recognition on degraded video streams, using the large NTU-RGB+D-120 dataset and five leading models. It applies three forms of temporal degradation — uniform subsampling, random subsampling, and single-block dropout — and shows that the form of degradation changes model rankings dramatically; at the same effective frame rate, accuracy can vary by more than 40% depending on how frames are removed. The paper identifies temporal regularity of the surviving frames as the main driver, and shows that a simple linear-interpolation mitigation recovers large accuracy losses for models that do not already resample inputs to a fixed length. Its headline result is that LogSigRNN, the worst model on clean data, becomes the best at 3 FPS, beating the state-of-the-art DeGCN by about 6% on average in five of six comparisons after interpolation mitigation. The paper argues that robustness evaluations restricted to random subsampling are misleading for real-world deployment.

What carries the argument

The central object is the benchmark itself: three degradation operators applied to NTU-RGB+D-120 test streams at matched effective dropout rates — uniform subsampling at 15/10/5/3 FPS, random subsampling at 0–90% dropout, and single contiguous block dropout at the same rates. The explanatory mechanism is the temporal regularity of surviving frames, and the model-level mechanism is the log-signature transform, a feature map built from iterated integrals of a path that is invariant to time reparametrisation, making it robust to variable or missing frames. The authors use linear interpolation of dropped frames as a simple inference-time mitigation, and they compare models with a 64-frame resampling preprocessing step (DeGCN, FR-HEAD, GCN-Dev-LSTM) against models that operate on 300 frames (MS-G3D Net, LogSigRNN).

What would settle it

Run all five models on identical retained frames at 3 FPS, feeding every model exactly the same set of, say, 64 uniformly subsampled frames (or otherwise matching input lengths), and compare accuracies. If LogSigRNN no longer beats DeGCN under this preprocessing-matched comparison, the headline ranking is an artifact of input-length differences rather than a property of the architecture.

Watch

Extended reading notes

Core claim

The central claim is that a model's rank in skeletal human action recognition is not fixed: it depends on the form of degradation applied to the test stream. On clean 30 FPS data, DeGCN leads with 87.2% cross-subject accuracy while LogSigRNN trails at 76.0%; after uniform subsampling to 3 FPS and linear-interpolation mitigation, LogSigRNN reaches 62.4% while DeGCN reaches 55.7%, and the same pattern holds for random subsampling. The authors locate the explanatory mechanism in the temporal regularity of retained frames, and they attribute LogSigRNN's resilience to the time-reparametrisation invariance of the log-signature transform from Rough Path Theory. They conclude that robustness assessment must vary the type of degradation, not just the rate, and that rough-path architectures deserve attention for low-frame-rate edge deployment.

Load-bearing premise

The ranking assumes that dropping the same fraction of raw frames before each model's own preprocessing gives the models informationally comparable test inputs; in reality the models consume different numbers of frames (64 for some, 300 for others), so LogSigRNN's low-frame-rate edge may come from that preprocessing mismatch rather than from a general degradation resistance.

Editorial extensions

If this is right

  • Benchmarks that only random-subsample frames can mis-rank models for real-world deployment; uniform and block-dropout degradation should become standard evaluation axes.
  • At low real-world frame rates (3–5 FPS), LogSigRNN should be preferred over DeGCN despite its lower clean-data accuracy.
  • Linear interpolation of dropped frames is a cheap, effective mitigation, recovering up to more than 40% accuracy in models without built-in fixed-length resampling.
  • Signature-based (Rough Path Theory) architectures are promising for degraded streaming input, but not all such architectures — for example GCN-Dev-LSTM — share this robustness.
  • The cross-subject and cross-setup train-test splits give qualitatively consistent results, supporting the stability of the ranking changes.

Reading between the lines

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

  • If the preprocessing mismatch is material, a fairer comparison that gives all models the same number of frames could shrink or reverse LogSigRNN's apparent advantage; this is a testable prediction the paper does not run.
  • The interpolation mitigation assumes the positions of dropped frames are known; in real edge pipelines the model often does not know which frames were lost, so the reported gains may overstate what is achievable without that side information.
  • The three degradation types are end-members of a regularity spectrum; real streams mix them (e.g., compression plus packet loss), so a natural extension is to benchmark mixtures and to train models with degradation augmentation.
  • The benchmark's methodology could transfer to other skeletal datasets and to other video classification tasks, where degradation type may also flip model rankings.
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

2 major / 4 minor

Summary. The paper introduces SHARDeg, a benchmark that measures the robustness of five skeletal human action recognition (SHAR) models on NTU-RGB+D-120 under three temporal degradation types: uniform subsampling, random subsampling, and single block dropout. The authors report that the form of degradation strongly affects accuracy, propose linear interpolation as an inference-time mitigation, and show that interpolation substantially improves some models at high dropout rates. Their headline claim is that LogSigRNN, the worst model on clean data, outperforms the current SoTA DeGCN in five of six comparisons at 3 FPS after mitigation. The paper supplies full accuracy tables for both cross-subject and cross-setup splits.

Significance. If the central claims hold, the benchmark provides a useful community resource: it demonstrates that degradation type, not just dropout volume, materially changes model rankings, and it makes a concrete, falsifiable prediction about LogSigRNN's low-frame-rate advantage. The paper is transparent in releasing complete data tables and a clear degradation protocol, and it validates its main qualitative findings across two train-test splits. However, the headline re-ranking claim rests on a comparison in which models receive different numbers of input frames, as the authors themselves partially acknowledge, and on single-run point estimates with no error bars. The benchmark's value as a resource is real, but the specific architectural conclusion about LogSigRNN needs stronger evidence.

major comments (2)
  1. [Section 2.2 / Contribution 3] The central claim that LogSigRNN is degradation-resistant and outperforms DeGCN at low frame rates is confounded by uncontrolled differences in input length. Section 2.2 states that DeGCN, FR-HEAD, and GCN-Dev-LSTM bilinearly resample inputs to 64 frames, while MS-G3D and LogSigRNN use 300-frame streams with padding or repetition. Because degradation is applied before these preprocessing steps, at 3 FPS with mitigation LogSigRNN sees a 300-frame interpolated sequence while DeGCN sees a 64-frame resampled version. The paper itself attributes the similar degradation curves of the three resampling models to this shared preprocessing, which is indirect evidence that preprocessing materially affects behavior. Yet Section 3 does not discuss the consequences of this asymmetry for the LogSigRNN-vs-DeGCN comparison; it merely notes the grouped curves. The 'five out of six' re-ranking therefore does not establish a general degradation-resistance property of the LogSigRNN architecture; it may be an artifact of input length. An input-matched experiment (e.g., feeding LogSigRNN the same 64-frame resampled input, or feeding DeGCN the 300-frame padded/interpolated stream) is required to support Contribution 3.
  2. [Supplementary Section 4 / Table 1] The headline 'five out of six cases' is based on single-run point estimates with no reported variance, seed variation, or statistical testing. The authors themselves state in Supplementary Section 4 that one of the six DeGCN-vs-LogSigRNN comparisons 'could be within error bounds' and that 'further tests over multiple random seeds would be required for clarification, which was beyond the computational resources available here.' That caveat applies equally to the claimed wins, several of which are small (e.g., 0.9 percentage points for single-block X-Sub in Table 1). For a benchmark whose main contribution is a ranking reversal, the absence of any uncertainty quantification is a load-bearing gap. Please provide at least repeated runs over multiple seeds for the key LogSigRNN-vs-DeGCN comparisons, or confidence intervals, so that readers can judge whether the reversal is real.
minor comments (4)
  1. [Abstract / Section 3] The abstract claims that model accuracy can vary by '>40%' depending on degradation type, but the reported DeGCN numbers at 90% effective dropout (55.3% uniform, 42.7% random, 16.3% block, from Tables 3, 5, 7) give a maximum spread of 39 percentage points. The main text says 'almost 40%.' The abstract wording should be aligned with the actual measurements.
  2. [Section 2.1] The description of excluding the first and last frames from 'all of the degradation experiments with and without mitigation' is ambiguous: does this exclusion also apply to the 0% dropout baseline? The baseline accuracies in Table 1 match the published values, so it appears the baseline is unaffected, but the text should state this explicitly.
  3. [Section 3] The claim that interpolation mitigation is a 'simple post-hoc mitigation' assumes the positions of dropped frames are known (stated in Section 2.1). This is clear, but the paper should also note that in real-world deployment, frame-drop metadata may not be available, which limits the direct applicability of the proposed mitigation.
  4. [Figure 1 / Figure 2] The figure captions reference a model key in panel (h), but in a two-column format this may be hard to read; consider a separate legend or repeated keys in the subfigures.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the benchmark's rankings and mitigation effects are empirical measurements, not derived from their own assumptions.

full rationale

SHARDeg is an empirical benchmark. The central claims—that degradation type changes accuracy by >40%, that interpolation mitigation improves some models, and that LogSigRNN overtakes DeGCN at 3 FPS—are measured outcomes from trained models evaluated on degraded NTU-RGB+D-120 test sets, not quantities fitted to or defined in terms of the conclusions. No model parameter is fitted to the degraded-test accuracies; baseline weights are either supplied by the original papers or trained on clean data, and the mitigated runs apply the same fixed linear-interpolation procedure to all models. The 'temporal regularity drives performance' explanation is a post-hoc interpretation of the measured ordering (uniform better than random better than block), not an input to the accuracy computation. The only self-citation ([49], Lo, Mital et al.) appears in a future-work suggestion and is not load-bearing. The differing internal frame counts (64-frame resampling for DeGCN/FR-HEAD/GCN-Dev-LSTM vs 300-frame streams for MS-G3D/LogSigRNN, Section 2.2) are a possible validity confound for the LogSigRNN comparison, and the paper itself discusses the consequences; that is a correctness/fairness concern, not circularity. No equation or definition reduces a claimed result to its own premise, so the circularity score is 0.

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

The benchmark adds no fitted constants or new entities; it depends on domain assumptions about test-time degradation, preprocessing fairness, and interpolation validity.

assumptions (3)
  • domain assumption Test-time degradation without retraining represents realistic deployment conditions.
    Section 2.3: models are trained on clean data and tested on degraded streams; the paper assumes real-world pipelines cannot retrain for every corruption.
  • domain assumption Different internal preprocessing across models does not bias the comparison at equal nominal dropout rate.
    Section 2.2: three models downsample to 64 frames, two use 300 frames; this asymmetry is acknowledged but treated as a model property, not a confound.
  • domain assumption Linear interpolation between remaining frames reconstructs missing skeleton positions adequately.
    Section 2.1: dropped frames are replaced by linear interpolation of adjacent available frames; used as the mitigation across all experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SHARDeg: A Benchmark for Skeletal Human Action Recognition in Degraded Scenarios." pith.science (2026). https://pith.science/paper/D7ACY7HJ

@misc{pith2026250518048,
  author       = {Pith},
  title        = {Pith review of: SHARDeg: A Benchmark for Skeletal Human Action Recognition in Degraded Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D7ACY7HJ}},
  note         = {Machine review of arXiv:2505.18048}
}
read the original abstract

Computer vision (CV) models for detection, prediction or classification tasks operate on video data-streams that are often degraded in the real world, due to deployment in real-time or on resource-constrained hardware. It is therefore critical that these models are robust to degraded data, but state of the art (SoTA) models are often insufficiently assessed with these real-world constraints in mind. This is exemplified by Skeletal Human Action Recognition (SHAR), which is critical in many CV pipelines operating in real-time and at the edge, but robustness to degraded data has previously only been shallowly and inconsistently assessed. Here we address this issue for SHAR by providing an important first data degradation benchmark on the most detailed and largest 3D open dataset, NTU-RGB+D-120, and assess the robustness of five leading SHAR models to three forms of degradation that represent real-world issues. We demonstrate the need for this benchmark by showing that the form of degradation, which has not previously been considered, has a large impact on model accuracy; at the same effective frame rate, model accuracy can vary by >40% depending on degradation type. We also identify that temporal regularity of frames in degraded SHAR data is likely a major driver of differences in model performance, and harness this to improve performance of existing models by up to >40%, through employing a simple mitigation approach based on interpolation. Finally, we highlight how our benchmark has helped identify an important degradation-resistant SHAR model based in Rough Path Theory; the LogSigRNN SHAR model outperforms the SoTA DeGCN model in five out of six cases at low frame rates by an average accuracy of 6%, despite trailing the SoTA model by 11-12% on un-degraded data at high frame rates (30 FPS).

Figures

Figures reproduced from arXiv: 2505.18048 by the authors.

Figure 1
Figure 1. Test accuracy versus effective dropout rate for three different experiments, Uniform [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Comparison of LogSigRNN (blue triangles) and DeGCN (green squares) models with [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Test accuracy versus effective dropout rate under the cross-setup train-test split for three [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 51 canonical work pages

  1. [1]

    Computer-Vision-Oriented Adaptive Sampling In Compressive Sensing,

    L. Liu, H. Nishikawa, J. Zhou, I. Taniguchi, and T. Onoye, “Computer-Vision-Oriented Adaptive Sampling In Compressive Sensing,”Sensors, vol. 24, no. 13, p. 4348, 2024

    work page 2024

  2. [2]

    Image And Video Compression With Neural Networks: A Review,

    S. Ma, X. Zhang, C. Jia, Z. Zhao, S. Wang, and S. Wang, “Image And Video Compression With Neural Networks: A Review,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 30, no. 6, pp. 1683–1698, 2020

    work page 2020

  3. [3]

    A Survey Of Methods For Low-Power Deep Learning And Computer Vision,

    A. Goel, C. Tung, Y .-H. Lu, and G. K. Thiruvathukal, “A Survey Of Methods For Low-Power Deep Learning And Computer Vision,” in2020 IEEE 6th World Forum on Internet of Things (WF-IoT), pp. 1–6, 2020

    work page 2020

  4. [4]

    Hardware Faults That Matter: Understanding And Estimating The Safety Impact Of Hardware Faults On Object Detection DNNs,

    S. Qutub, F. Geissler, Y . Peng, R. Gräfe, M. Paulitsch, G. Hinz, and A. Knoll, “Hardware Faults That Matter: Understanding And Estimating The Safety Impact Of Hardware Faults On Object Detection DNNs,” inInternational Conference on Computer Safety, Reliability, and Security, pp. 298–318, Springer, 2022

    work page 2022

  5. [5]

    Benchmarking Robustness In Object Detection: Autonomous Driving When Winter Is Coming,

    C. Michaelis, B. Mitzkus, R. Geirhos, E. Rusak, O. Bringmann, A. S. Ecker, M. Bethge, and W. Brendel, “Benchmarking Robustness In Object Detection: Autonomous Driving When Winter Is Coming,”arXiv preprint arXiv:1907.07484, 2019

    arXiv 1907

  6. [6]

    Benchmarking Neural Network Robustness To Common Corruptions And Perturbations,

    D. Hendrycks and T. Dietterich, “Benchmarking Neural Network Robustness To Common Corruptions And Perturbations,”arXiv preprint arXiv:1903.12261, 2019

    arXiv 1903

  7. [7]

    Action Recognition for Human–Robot Teaming: Exploring Mutual Performance Monitoring Possibilities,

    S. Mehak, J. D. Kelleher, M. Guilfoyle, and M. C. Leva, “Action Recognition for Human–Robot Teaming: Exploring Mutual Performance Monitoring Possibilities,”Machines, vol. 12, no. 1, 2024

    work page 2024

  8. [8]

    JOSENet: A Joint Stream Embedding Network for Violence Detection in Surveillance Videos

    P. Nardelli and D. Comminiello, “JOSENet: A Joint Stream Embedding Network for Violence Detection in Surveillance Videos,”arXiv preprint arXiv:2405.02961, 2024

    work page Pith review arXiv 2024

Show all 58 references
  1. [9]

    Elevating Urban Surveillance: A Deep CCTV Monitoring System for Detection of Anomalous Events via Human Action Recognition,

    H. Kim, H. Jeon, D. Kim, and J. Kim, “Elevating Urban Surveillance: A Deep CCTV Monitoring System for Detection of Anomalous Events via Human Action Recognition,”Sustainable Cities and Society, vol. 114, p. 105793, 2024

    2024

  2. [10]

    Human Activity Recognition (HAR) in Healthcare,

    L. Bibbò and M. M. B. R. Vellasco, “Human Activity Recognition (HAR) in Healthcare,”Applied Sciences, vol. 13, no. 24, 2023

    2023

  3. [11]

    Human Posture Estimation and Action Recognition on Fitness Behavior and Fitness,

    Y . Zhang, C. Zhao, Y . Yao, C. Wang, G. Cai, and G. Wang, “Human Posture Estimation and Action Recognition on Fitness Behavior and Fitness,”Alexandria Engineering Journal, vol. 107, pp. 434–442, 2024

    2024

  4. [12]

    OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields,

    Z. Cao, G. Hidalgo, T. Simon, S.-E. Wei, and Y . Sheikh, “OpenPose: Realtime Multi-Person 2D Pose Estimation Using Part Affinity Fields,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 43, no. 1, pp. 172–186, 2019. 10

    2019

  5. [13]

    CCTV Scene Perspective Distortion Estimation From Low-Level Motion Features,

    O. Arandjelovi´c, D.-S. Pham, and S. Venkatesh, “CCTV Scene Perspective Distortion Estimation From Low-Level Motion Features,”IEEE Transactions on Circuits and Systems for Video Technology, vol. 26, no. 5, pp. 939–949, 2016

    2016

  6. [14]

    Smart Frame Selection for Action Recognition,

    S. N. Gowda, M. Rohrbach, and L. Sevilla-Lara, “Smart Frame Selection for Action Recognition,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, pp. 1451–1459, 2021

    2021

  7. [15]

    Real-Time Human Action Recognition on Embedded Platforms,

    R. Wang, Z. Wang, P. Gao, M. Li, J. Jeong, Y . Xu, Y . Lee, C. M. Baum, L. T. Connor, and C. Lu, “Real-Time Human Action Recognition on Embedded Platforms,”arXiv preprint arXiv:2409.05662, 2024

    2024 arXiv

  8. [16]

    A Study of Frame Rate Optimisation for Action Clas- sification Inference,

    K. Anil, M. Bouroche, and B. Schoen-Phelan, “A Study of Frame Rate Optimisation for Action Clas- sification Inference,” in2024 9th International Conference on Image, Vision and Computing (ICIVC), pp. 215–219, IEEE, 2024

    2024

  9. [17]

    HMDB: A Large Video Database For Human Motion Recognition,

    H. Kuehne, H. Jhuang, E. Garrote, T. Poggio, and T. Serre, “HMDB: A Large Video Database For Human Motion Recognition,” in2011 International Conference on Computer Vision, pp. 2556–2563, 2011

    2011

  10. [18]

    The Kinetics Human Action Video Dataset,

    W. Kay, J. Carreira, K. Simonyan, B. Zhang, C. Hillier, S. Vijayanarasimhan, F. Viola, T. Green, T. Back, P. Natsev,et al., “The Kinetics Human Action Video Dataset,”arXiv preprint arXiv:1705.06950, 2017

    2017 arXiv

  11. [19]

    NTU RGB+D 120: A Large-Scale Benchmark for 3D Human Activity Understanding,

    J. Liu, A. Shahroudy, M. Perez, G. Wang, L.-Y . Duan, and A. C. Kot, “NTU RGB+D 120: A Large-Scale Benchmark for 3D Human Activity Understanding,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 42, no. 10, pp. 2684–2701, 2019

    2019

  12. [20]

    Investigating the Impact of Frame Rate Towards Robust Human Action Recognition,

    F. Harjanto, Z. Wang, S. Lu, A. C. Tsoi, and D. D. Feng, “Investigating the Impact of Frame Rate Towards Robust Human Action Recognition,”Signal Processing, vol. 124, pp. 220–232, 2016

    2016

  13. [21]

    On the Effects of Low Video Quality in Human Action Recognition,

    J. See and S. Rahman, “On the Effects of Low Video Quality in Human Action Recognition,” in2015 International Conference on Digital Image Computing: Techniques and Applications (DICTA), pp. 1–8, IEEE, 2015

    2015

  14. [22]

    Scsampler: Sampling Salient Clips from Video for Efficient Action Recognition,

    B. Korbar, D. Tran, and L. Torresani, “Scsampler: Sampling Salient Clips from Video for Efficient Action Recognition,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6232–6242, 2019

    2019

  15. [23]

    Multi-Agent Reinforcement Learning Based Frame Sampling for Effective Untrimmed Video Recognition,

    W. Wu, D. He, X. Tan, S. Chen, and S. Wen, “Multi-Agent Reinforcement Learning Based Frame Sampling for Effective Untrimmed Video Recognition,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 6222–6231, 2019

    2019

  16. [24]

    Faster Recurrent Networks for Efficient Video Classification,

    L. Zhu, D. Tran, L. Sevilla-Lara, Y . Yang, M. Feiszli, and H. Wang, “Faster Recurrent Networks for Efficient Video Classification,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 34, pp. 13098–13105, 2020

    2020

  17. [25]

    Developing the Path Signature Methodology and Its Application to Landmark-Based Human Action Recognition,

    W. Yang, T. Lyons, H. Ni, C. Schmid, and L. Jin, “Developing the Path Signature Methodology and Its Application to Landmark-Based Human Action Recognition,” inStochastic Analysis, Filtering, and Stochastic Optimization: A Commemorative Volume to Honor Mark HA Davis’s Contribut...

    2022

  18. [26]

    Logsig-RNN: A Novel Network for Robust and Efficient Skeleton-Based Action Recognition,

    S. Liao, T. Lyons, W. Yang, K. Schlegel, and H. Ni, “Logsig-RNN: A Novel Network for Robust and Efficient Skeleton-Based Action Recognition,”arXiv preprint arXiv:2110.13008, 2021

    2021 arXiv

  19. [27]

    Benchmarking The Robustness Of Spatial-Temporal Models Against Corruptions,

    C. Yi, S. Yang, H. Li, Y .-p. Tan, and A. Kot, “Benchmarking The Robustness Of Spatial-Temporal Models Against Corruptions,”arXiv preprint arXiv:2110.06513, 2021

    2021 arXiv

  20. [28]

    When Human Pose Estimation Meets Robustness: Adversarial Algorithms And Benchmarks,

    J. Wang, S. Jin, W. Liu, W. Liu, C. Qian, and P. Luo, “When Human Pose Estimation Meets Robustness: Adversarial Algorithms And Benchmarks,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11855–11864, 2021

    2021

  21. [29]

    Sequence of the Most Informative Joints (SMIJ): A New Representation for Human Skeletal Action Recognition,

    F. Ofli, R. Chaudhry, G. Kurillo, R. Vidal, and R. Bajcsy, “Sequence of the Most Informative Joints (SMIJ): A New Representation for Human Skeletal Action Recognition,”Journal of Visual Communication and Image Representation, vol. 25, no. 1, pp. 24–38, 2014

    2014

  22. [30]

    Human Action Recognition by Representing 3D Skeletons as Points in a Lie Group,

    R. Vemulapalli, F. Arrate, and R. Chellappa, “Human Action Recognition by Representing 3D Skeletons as Points in a Lie Group,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 588–595, 2014

    2014

  23. [31]

    RNN Fisher Vectors for Action Recognition and Image Annotation,

    G. Lev, G. Sadeh, B. Klein, and L. Wolf, “RNN Fisher Vectors for Action Recognition and Image Annotation,” inComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part VI, pp. 833–850, Springer, 2016

    2016

  24. [32]

    View Adaptive Recurrent Neural Networks for High Performance Human Action Recognition from Skeleton Data,

    P. Zhang, C. Lan, J. Xing, W. Zeng, J. Xue, and N. Zheng, “View Adaptive Recurrent Neural Networks for High Performance Human Action Recognition from Skeleton Data,” inProceedings of the IEEE International Conference on Computer Vision, pp. 2117–2126, 2017

    2017

  25. [33]

    Spatio-Temporal LSTM with Trust Gates for 3D Human Action Recognition,

    J. Liu, A. Shahroudy, D. Xu, and G. Wang, “Spatio-Temporal LSTM with Trust Gates for 3D Human Action Recognition,” inEuropean Conference on Computer Vision, pp. 816–833, Springer, 2016. 11

    2016

  26. [34]

    A Survey on 3D Skeleton-based Action Recognition Using Learning Method,

    B. Ren, M. Liu, R. Ding, and H. Liu, “A Survey on 3D Skeleton-based Action Recognition Using Learning Method,”Cyborg and Bionic Systems, vol. 5, p. 0100, 2024

    2024

  27. [35]

    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, 2018

    2018

  28. [36]

    Two-Stream Adaptive Graph Convolutional Networks for Skeleton- Based Action Recognition,

    L. Shi, Y . Zhang, J. Cheng, and H. Lu, “Two-Stream Adaptive Graph Convolutional Networks for Skeleton- Based Action Recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12026–12035, 2019

    2019

  29. [37]

    Skeleton-Based Action Recognition with Multi-Stream Adaptive Graph Convolutional Networks,

    L. Shi, Y . Zhang, J. Cheng, and H. Lu, “Skeleton-Based Action Recognition with Multi-Stream Adaptive Graph Convolutional Networks,”IEEE Transactions on Image Processing, vol. 29, pp. 9532–9545, 2020

    2020

  30. [38]

    Joint-Bone Fusion Graph Convolutional Network for Semi- Supervised Skeleton Action Recognition,

    Z. Tu, J. Zhang, H. Li, Y . Chen, and J. Yuan, “Joint-Bone Fusion Graph Convolutional Network for Semi- Supervised Skeleton Action Recognition,”IEEE Transactions on Multimedia, vol. 25, pp. 1819–1831, 2022

    2022

  31. [39]

    Skateformer: Skeletal-Temporal Transformer For Human Action Recognition,

    J. Do and M. Kim, “Skateformer: Skeletal-Temporal Transformer For Human Action Recognition,” in European Conference on Computer Vision, pp. 401–420, Springer, 2024

    2024

  32. [40]

    Disentangling and Unifying Graph Convolutions for Skeleton-Based Action Recognition,

    Z. Liu, H. Zhang, Z. Chen, Z. Wang, and W. Ouyang, “Disentangling and Unifying Graph Convolutions for Skeleton-Based Action Recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 143–152, 2020

    2020

  33. [41]

    Learning Discriminative Representations for Skeleton Based Action Recognition,

    H. Zhou, Q. Liu, and Y . Wang, “Learning Discriminative Representations for Skeleton Based Action Recognition,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 10608–10617, 2023

    2023

  34. [42]

    Degcn: Deformable Graph Convolutional Networks for Skeleton-Based Action Recognition,

    W. Myung, N. Su, J.-H. Xue, and G. Wang, “Degcn: Deformable Graph Convolutional Networks for Skeleton-Based Action Recognition,”IEEE Transactions on Image Processing, vol. 33, pp. 2477–2490, 2024

    2024

  35. [43]

    GCN-DEVLSTM: Path Development for Skeleton-Based Action Recognition,

    L. Jiang, W. Yang, X. Zhang, and H. Ni, “GCN-DEVLSTM: Path Development for Skeleton-Based Action Recognition,”arXiv Preprint arXiv:2403.15212, 2024

    2024 arXiv

  36. [44]

    Path Development Network with Finite-Dimensional Lie Group,

    H. Lou, S. Li, and H. Ni, “Path Development Network with Finite-Dimensional Lie Group,”Transactions on Machine Learning Research, 2024

    2024

  37. [45]

    Ntu RGB+D: A Large Scale Dataset for 3D Human Activity Analysis,

    A. Shahroudy, J. Liu, T.-T. Ng, and G. Wang, “Ntu RGB+D: A Large Scale Dataset for 3D Human Activity Analysis,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1010–1019, 2016

    2016

  38. [46]

    Channel-Wise Topology Refinement Graph Convolution for Skeleton-Based Action Recognition,

    Y . Chen, Z. Zhang, C. Yuan, B. Li, Y . Deng, and W. Hu, “Channel-Wise Topology Refinement Graph Convolution for Skeleton-Based Action Recognition,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 13359–13368, 2021

    2021

  39. [47]

    Neural Controlled Differential Equations For Irregular Time Series,

    P. Kidger, J. Morrill, J. Foster, and T. Lyons, “Neural Controlled Differential Equations For Irregular Time Series,”Advances in Neural Information Processing Systems, 2020

    2020

  40. [48]

    Neural Rough Differential Equations For Long Time Series,

    J. Morrill, C. Salvi, P. Kidger, J. Foster, and T. Lyons, “Neural Rough Differential Equations For Long Time Series,” inThirty-eighth International Conference on Machine Learning (ICML 2021), Proceedings of Machine Learning Research, pp. 7829–7838, Journal of Machine Learning ...

    2021

  41. [49]

    Collaborative Semantic Communication For Edge Inference,

    W. F. Lo, N. Mital, H. Wu, and D. Gündüz, “Collaborative Semantic Communication For Edge Inference,” IEEE Wireless Communications Letters, vol. 12, no. 7, pp. 1125–1129, 2023

    2023

  42. [50]

    Human Activity Recognition Under Partial Occlusion,

    I.-A. Kostis, E. Mathe, E. Spyrou, and P. Mylonas, “Human Activity Recognition Under Partial Occlusion,” inInternational Conference on Engineering Applications of Neural Networks, pp. 297–309, Springer, 2022

    2022

  43. [51]

    Evaluating 3D Human Pose Estimation In Occluded Multi-Sensor Scenarios: Dataset And Annotation Approach,

    K. Riou, K. Dong, Y . Huang, K. Subrin, D. P. Le Callet, and Y . Sun, “Evaluating 3D Human Pose Estimation In Occluded Multi-Sensor Scenarios: Dataset And Annotation Approach,” in2024 IEEE International Conference on Image Processing (ICIP), pp. 2683–2689, IEEE, 2024

    2024

  44. [52]

    SmoothNet: A Plug-And-Play Network For Refining Human Poses In Videos,

    A. Zeng, L. Yang, X. Ju, J. Li, J. Wang, and Q. Xu, “SmoothNet: A Plug-And-Play Network For Refining Human Poses In Videos,” inEuropean Conference on Computer Vision, Springer, 2022

    2022

  45. [53]

    T. J. Lyons, M. Caruana, and T. Lévy,Differential Equations Driven by Rough Paths. Springer, 2007

    2007

  46. [54]

    Rough Paths, Signatures and the Modelling of Functions on Streams,

    T. Lyons, “Rough Paths, Signatures and the Modelling of Functions on Streams,”arXiv Preprint arXiv:1405.4537, 2014

    2014 arXiv

  47. [55]

    A Primer on the Signature Method in Machine Learning,

    I. Chevyrev and A. Kormilitzin, “A Primer on the Signature Method in Machine Learning,”arXiv Preprint arXiv:1603.03788, 2016

    2016

  48. [56]

    LPSNet: A novel Log Path Signature Feature Based Hand Gesture Recognition Framework,

    C. Li, X. Zhang, and L. Jin, “LPSNet: A novel Log Path Signature Feature Based Hand Gesture Recognition Framework,” inProceedings of the IEEE International Conference on Computer Vision Workshops, pp. 631– 639, 2017. 12

    2017

  49. [57]

    Skeleton-Based Gesture Recognition Using Several Fully Connected Layers With Path Signature Features And Temporal Transformer Module,

    C. Li, X. Zhang, L. Liao, L. Jin, and W. Yang, “Skeleton-Based Gesture Recognition Using Several Fully Connected Layers With Path Signature Features And Temporal Transformer Module,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 33, pp. 8585–8593, 2019

    2019

  50. [58]

    Human Action Recognition In Unconstrained Trimmed Videos Using Residual Attention Network And Joints Path Signature,

    T. Ahmad, L. Jin, J. Feng, and G. Tang, “Human Action Recognition In Unconstrained Trimmed Videos Using Residual Attention Network And Joints Path Signature,”IEEE Access, vol. 7, pp. 121212–121222, 2019. 13 Supplementary Material 1 Review of Rough Path Theory for Skeletal Huma...

    2019

Pith tools

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