Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Towards Efficient General Feature Prediction in Masked Skeleton Modeling

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that masked skeleton pretraining becomes faster and better when the prediction target changes from raw joint coordinates to a small hierarchy of semantic features—local motion over 5, 10, and 30 frames plus a global action

desk verdict A solid within-subfield advance: hierarchical high-level feature prediction makes masked skeleton pretraining much faster with strong results, but the anti-collapse mechanism at the center of the method is never directly tested. read the letter →

arxiv 2509.03609 v1 pith:XULPXR4J submitted 2025-09-03 cs.CV

Shengkai Sun , Zefan Zhang , Jianfeng Dong , Zhiyong Cheng , Xiaojun Chang , Meng Wang This is my paper · ORCID
classification cs.CV
keywords maskedskeletonmodelingself-supervisedactionrecognitionhigh-levelfeaturepredictiontargetgenerationnetworkvariance-covarianceregularizationretrievaltrainingefficiencyhierarchicalrepresentationlearning
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

Masked autoencoders for skeleton action recognition have mostly been taught to reconstruct missing joint coordinates, which is expensive and leaves a gap between pretraining and the semantics used downstream. This paper tries to close that gap by changing what gets predicted: instead of many low-level targets, the model predicts a compact hierarchy of high-level features, from short local motion patterns to a global action-level representation. Those targets are generated on the fly by a lightweight target network, with a variance-covariance penalty to keep the targets from collapsing to constants. If the claim holds, the practical payoff is concrete: pretraining runs 6.2 times faster on the same GPU (3h14m vs 20h27m), while linear-probe accuracy on NTU-60 x-view rises to 92.0 and retrieval on x-sub jumps from 62.0 to 70.9.

What carries the argument

The carrying object is a hierarchy of prediction targets built by cascaded temporal pooling. The decoder first predicts local joint motion over a 5-frame span, then progressively coarser spans of 10 and 30 frames, and finally a global action-level representation after the temporal dimension collapses to a single step; each level is supervised by a matching target from a lightweight MLP-based target generation network. The generation network is fed a simple motion-difference transform rather than raw or masked joints, and variance-covariance regularization keeps the batch-wise target features diverse and decorrelated. This design is what makes the target set small enough to cut decoder cost w

What would settle it

Remove the variance and covariance regularization terms from Eq. (6), rerun pretraining with identical hyperparameters, and measure the batch-wise variance of the generated target features over training; if the variance collapses toward zero while linear-probe accuracy on NTU-60 stays near chance, the joint-optimization stability claim is false.

Watch

Extended reading notes

Core claim

The paper's central claim is that the learning objective, not the encoder, is the main lever in masked skeleton modeling. Its proposed GFP replaces the standard reconstruction of masked joint coordinates with prediction of hierarchical high-level features: local motion patterns pooled over 5, 10, and 30 frames, and one global action-semantic vector. A lightweight target generation network produces these targets online from a motion-difference transformation of the input, so no precomputed offline features or bulky momentum encoder are needed; variance and covariance constraints on the target features prevent both networks from collapsing to constant outputs. The authors report that this sche

Load-bearing premise

The load-bearing premise is that the small jointly trained target network, constrained only by a diversity penalty on its outputs, keeps producing informative targets instead of collapsing to constant vectors; if that penalty fails, the decoder is predicting constants and the pretraining signal vanishes.

Editorial extensions

If this is right

  • Pretraining can run on a single consumer GPU in about three hours, so masked skeleton modeling becomes accessible without large compute budgets.
  • Recognition and retrieval both improve, suggesting the learned representation carries global semantics that low-level reconstruction misses.
  • The framework removes the need for offline or precomputed target features, making the pipeline simpler to apply to new skeleton datasets.
  • Ablations show all hierarchy levels contribute; dropping the global or local targets lowers accuracy, so target granularity matters independently of architecture.
  • In 1%- and 10%-labeled fine-tuning, the frozen encoder outperforms prior masked models, indicating better label efficiency.

Reading between the lines

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

  • If the speedup is mostly from the reduced target count and shallow decoder, the same 'predict fewer, higher-level targets' recipe could transfer to masked image or video modeling, where decoder cost is likewise a bottleneck; this is an extension the paper does not test.
  • The stability of the joint optimization is asserted rather than demonstrated directly; a natural experiment is removing the variance-covariance terms and tracking target-feature variance, since the method's non-collapse claim rests on that regularizer.
  • The motion-difference input works better than raw or masked joints, which hints that higher-order temporal derivatives or velocity-aware inputs could push target quality further.
  • Retrieval benefits more than recognition from the global branch, so ablating the global target on retrieval specifically would clarify where the semantic gain comes from; the paper reports recognition ablations only.
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 proposes GFP, a self-supervised masked skeleton modeling framework. Instead of reconstructing raw joint coordinates or motion trajectories, GFP trains an encoder-decoder to predict hierarchical high-level features: local features over 5-, 10-, and 30-frame windows plus one global semantic feature. These targets are produced online by a lightweight target generation network (TGN) that takes a temporal-difference transform of the full skeleton sequence. To prevent collapse, VICReg-style variance and covariance regularization is applied to the target features. The joint decoder/TGN objective is minimized during pretraining, and the frozen encoder is evaluated by linear probing, semi-supervised fine-tuning, and retrieval on NTU-60, NTU-120, and PKU-MMD II. The paper reports a 6.2x pretraining speedup over SkeletonMAE/MAMP and strong downstream numbers, e.g., NTU-60 x-view 92.0, PKU-MMD II 56.2, and retrieval x-sub 70.9.

Significance. If the central mechanism is validated, the paper makes a strong practical and conceptual contribution: it shows that masked skeleton modeling can be made much cheaper by predicting a small set of high-level semantic targets instead of hundreds of low-level coordinate targets, and that such targets can be generated online without a large EMA network. The paper's strengths are the breadth of downstream evaluation, the direct comparison of pretraining time on the same GPU, and the ablations of target hierarchy, TGN input, and TGN/projector architecture. However, the core claim—that the collaborative optimization converges to non-trivial, semantically meaningful targets—is not directly tested. The paper omits the key 'without Lreg' ablation and does not show loss curves or target feature statistics. The reported gains are therefore not yet attributable to high-level feature prediction, which is the paper's central contribution.

major comments (3)
  1. [Section 3.3, Eq. (7)] The paper does not state which parameters receive gradients from Lpred. If the TGN is updated by Lpred, as 'collaborative learning' implies, the objective has degenerate solutions: the TGN can output features predictable from the decoder, driving Lpred to zero while Lreg is satisfied by variance above gamma and near-zero covariance. Eqs. (4)-(6) enforce statistical diversity, not semantic content. No ablation removes Lreg, no training curves of Lpred or target variance are shown, and no stop-gradient/EMA is used. Thus the claim that variance-covariance regularization stabilizes the mutual guidance is empirically unsupported. Please add a w/o-Lreg ablation, report Lpred/target variance over epochs, and clarify or alter gradient flow (e.g., detach targets or use an EMA).
  2. [Section 4.4, Figure 3] The {5,10,30,g}* setting removes the TGN and switches to flattened low-level vectors, changing two variables at once: target representation and training signal. The drop from 92.0 to 86.8 therefore does not cleanly show the benefit of the TGN's online high-level targets. A controlled experiment would keep the decoder and target dimensions fixed and compare jointly trained TGN targets, frozen/offline TGN targets, random fixed targets, and direct low-level prediction. Without this, the contribution of the collaborative mechanism is confounded.
  3. [Section 3.3 and Section 4.2] The TGN and target feature construction are underspecified. Local extractors 'merge across consecutive frames' and the global extractor 'integrates all spatiotemporal context', but the exact operations, tensor shapes, and how Xt1 in R^{(Te/t1)*V*(t1*l*C)} maps to Zt1 are not given. The target dimension Ct is never stated; M=251 is reported but not derived. Since the nature of the targets is central to the method, full precision is needed for reproducibility. Please add pseudo-code or an algorithm block.
minor comments (5)
  1. [Section 4.5, Table 7] The text says experiments use 1% and 5% labeled data, but the table columns are 1% and 10%. Correct the discrepancy and state the exact protocol.
  2. [Section 3.3, Eq. (7)] Typo 'paramters'; also lambda, alpha, beta values appear only in Section 4.2, not at the equation. Please clarify which parameters are tuned.
  3. [Section 3.3] Xe = Xe[1:] - Xe[:-1] reuses Xe for the transformed tensor. Clarify boundary handling and use distinct notation for the motion transform.
  4. [Section 4.2] The pretraining setup describes the motion-aware masking strategy of [20] but does not state the mask ratio used in GFP. The introduction's 90% figure should be confirmed or corrected.
  5. [Table 2 / Abstract] On NTU-120 x-sub, GFP (79.1) is not the best (S-JEPA 79.6). The abstract's 'state-of-the-art' should be qualified per protocol or dataset.

Circularity Check

1 steps flagged · score 4.0 of 10

Pretraining supervision is self-referential: TGN-generated targets are jointly optimized against the same prediction loss, so 'high-level feature prediction' is not anchored to an external signal; downstream evaluation remains independent.

  1. self definitional [Section 3.3 (Target Generation / Optimization Objectives), Eqs. (3) and (7)]
    "This supervisory mechanism operates bidirectionally: The decoder infers high-level semantic features from unmasked joint features for learning, while the target generation network uses the decoder’s output as an optimization target... The collaborative optimization objective for both networks combines the prediction loss with information-maximization constraints through weighted aggregation: Ltotal = λLpred + Lreg"

    In Eq. (3), Lp = 1/M Σ ||Zpj − Ztj||^2_F, with Ztj generated by TGN. Eq. (7) states this same Lpred is part of the objective for 'both networks'; therefore the TGN parameters are also updated to minimize the discrepancy between Ztj and Zpj. At the optimum of the MSE term, the target Ztj can equal the decoder's prediction Zpj, so the 'prediction' is not anchored to any fixed external semantic signal—the target is defined by the co-optimization. The variance/covariance regularizer (Eqs. 4–6) only enforces statistical diversity; it does not supply semantic content. Thus the claim that the decoder learns high-level semantic features from TGN targets is self-referential by construction: the supervision is the decoder's own output projected back. The downstream linear-probe/retrieval evaluations

full rationale

The only potential circularity is in the pretraining target generation. I quote Eq. (3) and Eq. (7) to show that the TGN is optimized on the same Lpred that the decoder minimizes, making the 'high-level features' endogenous to the prediction loop. This fits the self_definitional pattern. However, the paper's headline results—6.2x speedup, NTU-60 x-view 92.0, PKU-MMD II 56.2, retrieval gains—are measured against fixed baselines and held-out labels, so they do not reduce to the pretraining objective. No load-bearing self-citations or imported uniqueness theorems were found; citations to VICReg [2], MAMP [20], S-JEPA [1] are external related work. The absence of an Lreg ablation or training curves is a correctness risk (the non-triviality of the mutual guidance is asserted, not demonstrated) but not itself circularity. Overall score 4 reflects partial circularity in the training objective with independent downstream validation.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several hand-chosen hyperparameters (λ, α, β, hierarchy windows, patch length) and on two unproven domain assumptions: that the frame-difference transform preserves semantics, and that VICReg-style regularization suffices to prevent collapse in the joint decoder-TGN optimization. The hierarchy choice is paper-specific and only empirically validated. No new physical or conceptual entities are introduced.

free parameters (7)
  • lambda (Lpred weight) = 5
    Trade-off between prediction loss and regularization in Ltotal (Eq. 7); chosen by authors without sensitivity analysis.
  • alpha (covariance weight) = 5
    Weight of covariance regularization term in Eq. 6; follows VICReg practice, no ablation.
  • beta (variance weight) = 1
    Weight of variance regularization term in Eq. 6; set to 1 following VICReg.
  • hierarchy windows t1,t2,t3 = 5,10,30 frames
    Temporal receptive fields for local feature targets; chosen experimentally (Fig. 3 investigates their impact).
  • patch length l = 4 frames
    Temporal patchification length in tokenization; standard choice from prior work.
  • target dimensionality = not specified
    Dimensions of Ztj / Zpg are not given; they must be matched by projectors, affecting capacity.
  • masking ratio = unstated
    The masking ratio is not reported; likely 90% following MAMP, but unverified.
assumptions (5)
  • domain assumption The transformation Xe' = Xe[1:] - Xe[:-1] preserves high-level motion semantics while removing bias toward frequently unmasked joints.
    Invoked in Section 3.3 to justify the TGN input; the paper provides no analysis that this holds beyond empirical accuracy.
  • domain assumption VICReg-style variance-covariance regularization prevents collapse and yields a non-trivial stable state in joint decoder-TGN optimization.
    Inspired by [2], but the paper does not ablate Lreg or show collapse without it; this is an assumption taken from the VICReg literature.
  • ad hoc to paper Predicting features at 5,10,30-frame windows plus a global representation is sufficient and complementary for skeleton action representation.
    The choice of hierarchy is paper-specific and validated only by ablation on NTU-60 x-view (Fig. 3), not by theory.
  • ad hoc to paper The joint optimization converges to a semantically consistent stable state (not a collapsed solution).
    Stated in Section 3.3 ('iterative optimization that converges to a semantically consistent stable state') but never directly measured; the regularizer is asserted to ensure it.
  • domain assumption Linear probing of the frozen encoder measures representation quality.
    Standard assumption in SSL evaluation, used throughout the experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Efficient General Feature Prediction in Masked Skeleton Modeling." pith.science (2026). https://pith.science/paper/XULPXR4J

@misc{pith2026250903609,
  author       = {Pith},
  title        = {Pith review of: Towards Efficient General Feature Prediction in Masked Skeleton Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XULPXR4J}},
  note         = {Machine review of arXiv:2509.03609}
}
abstract

Recent advances in the masked autoencoder (MAE) paradigm have significantly propelled self-supervised skeleton-based action recognition. However, most existing approaches limit reconstruction targets to raw joint coordinates or their simple variants, resulting in computational redundancy and limited semantic representation. To address this, we propose a novel General Feature Prediction framework (GFP) for efficient mask skeleton modeling. Our key innovation is replacing conventional low-level reconstruction with high-level feature prediction that spans from local motion patterns to global semantic representations. Specifically, we introduce a collaborative learning framework where a lightweight target generation network dynamically produces diversified supervision signals across spatial-temporal hierarchies, avoiding reliance on pre-computed offline features. The framework incorporates constrained optimization to ensure feature diversity while preventing model collapse. Experiments on NTU RGB+D 60, NTU RGB+D 120 and PKU-MMD demonstrate the benefits of our approach: Computational efficiency (with 6.2$\times$ faster training than standard masked skeleton modeling methods) and superior representation quality, achieving state-of-the-art performance in various downstream tasks.

Figures

Figures reproduced from arXiv: 2509.03609 by the authors.

Figure 1
Figure 1. Frameworks for general feature prediction (GFP). In [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The framework of our proposed GFP. The GFP framework establishes a bidirectional learning paradigm between two core [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Performance of various methods with different targets. [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

44 extracted references · 40 canonical work pages

  1. [34]

    Skeletonmae: Spatial-temporal masked au- toencoders for self-supervised skeleton action recognition

    Wenhan Wu, Yilei Hua, Ce Zheng, Shiqian Wu, Chen Chen, and Aidong Lu. Skeletonmae: Spatial-temporal masked au- toencoders for self-supervised skeleton action recognition. In 2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 224–229. IEEE, 2023. 3, 6, 7, 8

    work page 2023

  2. [39]

    Self-supervised 3d action representa- tion learning with skeleton cloud colorization

    Siyuan Yang, Jun Liu, Shijian Lu, Er Meng Hwa, Yongjian Hu, and Alex C Kot. Self-supervised 3d action representa- tion learning with skeleton cloud colorization. IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 2024. 6

    work page 2024

  3. [1]

    S-jepa: A joint embedding predictive architecture for skeletal action recog- nition

    Mohamed Abdelfattah and Alexandre Alahi. S-jepa: A joint embedding predictive architecture for skeletal action recog- nition. In European Conference on Computer Vision, pages 367–384. Springer, 2024. 3, 6, 7, 8

    work page 2024

  4. [2]

    Vi- creg: Variance-invariance-covariance regularization for self- supervised learning

    Adrien Bardes, Jean Ponce, and Yann Lecun. Vi- creg: Variance-invariance-covariance regularization for self- supervised learning. In ICLR 2022-International Conference on Learning Representations, 2022. 2, 5

    work page 2022

  5. [3]

    Hi- erarchically self-supervised transformer for human skeleton representation learning

    Yuxiao Chen, Long Zhao, Jianbo Yuan, Yu Tian, Zhaoyang Xia, Shijie Geng, Ligong Han, and Dimitris N Metaxas. Hi- erarchically self-supervised transformer for human skeleton representation learning. In European Conference on Com- puter Vision, pages 185–202, 2022. 8

    work page 2022

  6. [4]

    Skeleton-based action recognition with shift graph convolutional network

    Ke Cheng, Yifan Zhang, Xiangyu He, Weihan Chen, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with shift graph convolutional network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 183–192, 2020. 1

    work page 2020

  7. [5]

    Hierarchical transformer: Unsupervised representation learning for skeleton-based hu- man action recognition

    Yi-Bin Cheng, Xipeng Chen, Junhong Chen, Pengxu Wei, Dongyu Zhang, and Liang Lin. Hierarchical transformer: Unsupervised representation learning for skeleton-based hu- man action recognition. In 2021 IEEE International Con- ference on Multimedia and Expo (ICME) , pages 1–6, 2021. 3

    work page 2021

  8. [6]

    Hierarchical contrast for unsu- pervised skeleton-based action representation learning

    Jianfeng Dong, Shengkai Sun, Zhonglin Liu, Shujie Chen, Baolong Liu, and Xun Wang. Hierarchical contrast for unsu- pervised skeleton-based action representation learning. In Proceedings of the AAAI Conference on Artificial Intelli- gence, pages 525–533, 2023. 1, 6, 7, 8

    work page 2023

Show all 44 references
  1. [7]

    An image is worth 16x16 words: Trans- formers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale. arXiv preprint a...

    2010 arXiv

  2. [8]

    Slowfast networks for video recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202–6211, 2019. 1

    2019

  3. [9]

    Hyperbolic self-paced learning for self-supervised skeleton-based action representations

    Luca Franco, Paolo Mandica, Bharti Munjal, and Fabio Galasso. Hyperbolic self-paced learning for self-supervised skeleton-based action representations. arXiv preprint arXiv:2303.06242, 2023. 8

    2023 arXiv

  4. [10]

    Part aware contrastive learning for self-supervised action recognition

    Yilei Hua, Wenhan Wu, Ce Zheng, Aidong Lu, Mengyuan Liu, Chen Chen, and Shiqian Wu. Part aware contrastive learning for self-supervised action recognition. In Interna- tional Joint Conference on Artificial Intelligence , 2023. 3, 6

    2023

  5. [11]

    3d convolu- tional neural networks for human action recognition

    Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3d convolu- tional neural networks for human action recognition. IEEE transactions on pattern analysis and machine intelligence , 35(1):221–231, 2012. 1

    2012

  6. [12]

    Global-local motion transformer for unsupervised skeleton-based action learning

    Boeun Kim, Hyung Jin Chang, Jungho Kim, and Jin Young Choi. Global-local motion transformer for unsupervised skeleton-based action learning. In European Conference on Computer Vision, pages 209–225, 2022. 3, 8

    2022

  7. [13]

    Hmdb: a large video database for human motion recognition

    Hildegard Kuehne, Hueihan Jhuang, Est ´ıbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In 2011 Inter- national conference on computer vision , pages 2556–2563. IEEE, 2011. 1

    2011

  8. [14]

    Unsupervised feature learning of human actions as trajectories in pose embedding manifold

    Jogendra Nath Kundu, Maharshi Gor, Phani Krishna Uppala, and Venkatesh Babu Radhakrishnan. Unsupervised feature learning of human actions as trajectories in pose embedding manifold. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1459–1467, 2019. 2

    2019

  9. [15]

    3d human action rep- resentation learning via cross-view consistency pursuit

    Linguo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action rep- resentation learning via cross-view consistency pursuit. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 4741–4750, 2021. 3

    2021

  10. [16]

    Actionlet- dependent contrastive learning for unsupervised skeleton- based action recognition

    Lilang Lin, Jiahang Zhang, and Jiaying Liu. Actionlet- dependent contrastive learning for unsupervised skeleton- based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2363–2372, 2023. 3

    2023

  11. [17]

    Ntu rgb+d 120: A large- scale benchmark for 3d human activity understanding

    Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Ntu rgb+d 120: A large- scale benchmark for 3d human activity understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(10):2684–2701, 2019. 1, 5

    2019

  12. [18]

    A benchmark dataset and comparison study for multi-modal human action analytics

    Jiaying Liu, Sijie Song, Chunhui Liu, Yanghao Li, and Yueyu Hu. A benchmark dataset and comparison study for multi-modal human action analytics. ACM Transactions on Multimedia Computing, Communications, and Applications (TOMM), 16(2):1–24, 2020. 5

    2020

  13. [19]

    Cmd: Self-supervised 3d action representation learning with cross-modal mutual distillation

    Yunyao Mao, Wengang Zhou, Zhenbo Lu, Jiajun Deng, and Houqiang Li. Cmd: Self-supervised 3d action representation learning with cross-modal mutual distillation. In European Conference on Computer Vision , pages 734–752. Springer,

  14. [20]

    Masked motion predictors are strong 3d action representation learners

    Yunyao Mao, Jiajun Deng, Wengang Zhou, Yao Fang, Wanli Ouyang, and Houqiang Li. Masked motion predictors are strong 3d action representation learners. In Proceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 10181–10191, 2023. 3, 6, 7, 8

    2023

  15. [21]

    Unsupervised 3d human pose representation with viewpoint and pose disen- tanglement

    Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disen- tanglement. In European Conference on Computer Vision , pages 102–118, 2020. 2

    2020

  16. [22]

    Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition

    Haocong Rao, Shihao Xu, Xiping Hu, Jun Cheng, and Bin Hu. Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition. Information Sciences, 569:90–109, 2021. 3

    2021

  17. [23]

    Halp: Hal- lucinating latent positives for skeleton-based self-supervised learning of actions

    Anshul Shah, Aniket Roy, Ketul Shah, Shlok Mishra, David Jacobs, Anoop Cherian, and Rama Chellappa. Halp: Hal- lucinating latent positives for skeleton-based self-supervised learning of actions. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recogni...

    2023

  18. [24]

    Ntu rgb+d: A large scale dataset for 3d human activity anal- ysis

    Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+d: A large scale dataset for 3d human activity anal- ysis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1010–1019, 2016. 1, 5

    2016

  19. [25]

    Skeleton-based action recognition with directed graph neu- ral networks

    Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Skeleton-based action recognition with directed graph neu- ral networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 7912– 7921, 2019. 1

    2019

  20. [26]

    Two- stream adaptive graph convolutional networks for skeleton- based action recognition

    Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. Two- stream adaptive graph convolutional networks for skeleton- based action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 12026–12035, 2019. 1

    2019

  21. [27]

    Adversarial self-supervised learn- ing for semi-supervised 3d action recognition

    Chenyang Si, Xuecheng Nie, Wei Wang, Liang Wang, Tie- niu Tan, and Jiashi Feng. Adversarial self-supervised learn- ing for semi-supervised 3d action recognition. In European Conference on Computer Vision, pages 35–51, 2020. 8

    2020

  22. [28]

    Ucf101: A dataset of 101 human actions classes from videos in the wild

    Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 1

    2012 arXiv

  23. [29]

    Predict & cluster: Unsupervised skeleton based action recognition

    Kun Su, Xiulong Liu, and Eli Shlizerman. Predict & cluster: Unsupervised skeleton based action recognition. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9631–9640, 2020. 2, 6

    2020

  24. [30]

    Uni- fied multi-modal unsupervised representation learning for skeleton-based action understanding

    Shengkai Sun, Daizong Liu, Jianfeng Dong, Xiaoye Qu, Junyu Gao, Xun Yang, Xun Wang, and Meng Wang. Uni- fied multi-modal unsupervised representation learning for skeleton-based action understanding. In Proceedings of the 31st ACM International Conference on Multimedia , pages ...

    2023

  25. [31]

    Skeleton-contrastive 3d action representation learn- ing

    Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learn- ing. In Proceedings of the 29th ACM International Confer- ence on Multimedia, pages 1655–1663, 2021. 3, 6, 7, 8

    2021

  26. [32]

    Eulermormer: Robust eulerian motion magnification via dynamic filtering within transformer

    Fei Wang, Dan Guo, Kun Li, and Meng Wang. Eulermormer: Robust eulerian motion magnification via dynamic filtering within transformer. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5345–5353, 2024. 2

    2024

  27. [33]

    Usdrl: Unified skeleton-based dense repre- sentation learning with multi-grained feature decorrelation

    Wanjiang Weng, Hongsong Wang, Junbo He, Lei He, and Guosen Xie. Usdrl: Unified skeleton-based dense repre- sentation learning with multi-grained feature decorrelation. arXiv preprint arXiv:2412.09220, 2024. 6, 8

    2024 arXiv

  28. [35]

    Deep image clustering with contrastive learn- ing and multi-scale graph convolutional networks

    Yuankun Xu, Dong Huang, Chang-Dong Wang, and Jian- Huang Lai. Deep image clustering with contrastive learn- ing and multi-scale graph convolutional networks. Pattern Recognition, 2024. 1

    2024

  29. [36]

    Spatial tempo- ral graph convolutional networks for skeleton-based action recognition

    Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial tempo- ral graph convolutional networks for skeleton-based action recognition. In Thirty-second AAAI Conference on Artificial Intelligence, 2018. 1

    2018

  30. [37]

    View- invariant skeleton action representation learning via motion retargeting

    Di Yang, Yaohui Wang, Antitza Dantcheva, Lorenzo Garat- toni, Gianpiero Francesca, and Franc ¸ois Br ´emond. View- invariant skeleton action representation learning via motion retargeting. International Journal of Computer Vision, pages 1–16, 2024. 3

    2024

  31. [38]

    Skeleton cloud colorization for unsupervised 3d action representation learning

    Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C Kot. Skeleton cloud colorization for unsupervised 3d action representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13423– 13433, 2021. 2, 8

    2021

  32. [40]

    Contrastive positive mining for unsupervised 3d action representation learning

    Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wan- qing Li. Contrastive positive mining for unsupervised 3d action representation learning. In European Conference on Computer Vision, pages 36–51. Springer, 2022. 3, 8

    2022

  33. [41]

    Prompted con- trast with masked motion modeling: Towards versatile 3d ac- tion representation learning

    Jiahang Zhang, Lilang Lin, and Jiaying Liu. Prompted con- trast with masked motion modeling: Towards versatile 3d ac- tion representation learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7175–7183,

  34. [42]

    Tuber: Tubelet transformer for video action detection

    Jiaojiao Zhao, Yanyi Zhang, Xinyu Li, Hao Chen, Bing Shuai, Mingze Xu, Chunhui Liu, Kaustav Kundu, Yuanjun Xiong, Davide Modolo, et al. Tuber: Tubelet transformer for video action detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ,...

    2022

  35. [43]

    Temporal action detection with structured segment networks

    Yue Zhao, Yuanjun Xiong, Limin Wang, Zhirong Wu, Xi- aoou Tang, and Dahua Lin. Temporal action detection with structured segment networks. In Proceedings of the IEEE international conference on computer vision , pages 2914– 2923, 2017. 1

    2017

  36. [44]

    Unsupervised representation learning with long-term dynamics for skeleton based action recognition

    Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jian- hua Dai, and Zhefeng Gong. Unsupervised representation learning with long-term dynamics for skeleton based action recognition. In Proceedings of the AAAI Conference on Ar- tificial Intelligence, pages 2644–2651, 2018. 1, 2, 6

    2018

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.