Pith. sign in

REVIEW 4 major objections 6 minor 52 references

Spatio-Temporal Joint Density Driven Learning for Skeleton-Based Action Recognition

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

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

Pith's one-line read The paper claims that tracking a learnable spatio-temporal joint density to select 'prime joints' improves self-supervised skeleton-based action recognition, outperforming prior contrastive methods by up to 3.6 percentage points on NTU…

desk verdict The STJD idea is a genuine step forward, but the two principal tables give different numbers for the headline NTU120 X-sub result, so the main claim needs verification before this is citable. read the letter →

arxiv 2505.23012 v1 pith:DZIS5KI4 submitted 2025-05-29 cs.CV

Shanaka Ramesh Gunasekara , Wanqing Li , Philip Ogunbona , Jack Yang This is my paper
classification cs.CV
keywords skeleton-basedactionrecognitionself-supervisedlearningcontrastivekerneldensityestimationprimejointsspatio-temporaljointmaskedreconstructionmotionanalysis
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

Self-supervised skeleton-based action recognition has mostly focused on moving joints, but the paper argues that the interaction between moving joints and nearby static joints carries discriminative information that prior methods leave unused. It introduces the spatio-temporal joint density (STJD), a learnable kernel-density measure of how much each joint is influenced by the others at each frame, and uses its temporal change to detect a set of 'prime joints' that combine moving and static discriminative joints. These prime joints then steer two learning frameworks: STJD-CL, a contrastive method that aligns a whole sequence with its prime joints while repelling non-prime joints, and STJD-MP, a masked-reconstruction method that preferentially masks prime joints. On the NTU RGB+D 120 benchmark, STJD-CL outperforms the Actionlet-based ActCLR by 3.5 and 3.6 percentage points under cross-subject and cross-setup protocols, and STJD-MP reaches a new state of the art on the cross-setup protocol. The point of the work is that a data-driven, unsupervised notion of 'which joints matter' can improve both major families of self-supervised action learning.

What carries the argument

The central object is the spatio-temporal joint density $D_t(r)$, a kernel-density estimate of the influence of all joints on joint $r$ at frame $t$, computed with a Gaussian kernel and a sample-wise learnable bandwidth $h_i$ shared across channels. The load-bearing quantity is its temporal change $\Delta D_t(r) = D_{t+\delta t}(r) - D_t(r)$; after a Taylor-expansion argument, the paper interprets this change as combining each joint's own motion with its interaction with other joints. Thresholding $\Delta D_t(r)$ at a fixed $\beta$ (empirically set to 0.65) yields the prime joints, which are used as a mask to align, contrast, or reconstruct representations in STJD-CL and STJD-MP.

What would settle it

A concrete test: in STJD-CL, replace the STJD prime-joint mask with a random joint mask of the same average size, keeping all other training choices fixed; if downstream recognition accuracy does not fall, the density-based selection is not the cause of the reported gains. A complementary check is to inspect the detected prime joints on actions the paper says lose accuracy, such as playing with a mobile phone or typing, and test whether the fixed threshold systematically omits subtle hand joints on those sequences.

Watch

Extended reading notes

Core claim

The paper's central claim is that a spatio-temporal joint density, defined as a learnable Gaussian kernel-density estimate over the joints of a skeleton sequence in embedding space, can quantify the interaction between moving and static joints, and that tracking its temporal change identifies 'prime joints'—the action-discriminative subset not confined to predefined body parts. This detection is unsupervised and learnable, unlike earlier heuristics that pick the most moving joints or supervise joint importance with class labels. The authors show that using prime joints to guide contrastive learning (STJD-CL) and masked reconstruction (STJD-MP) consistently improves over their respective baselines, with the largest gains on the larger NTU RGB+D 120 dataset, and they report statistically significant improvement over ActCLR via paired t-tests.

Load-bearing premise

The load-bearing premise is that the temporal change of the learnable joint-density measure, evaluated in the embedding space and compared with a single fixed threshold, reliably identifies the joints that actually discriminate an action across different actions, subjects, views, and datasets.

Editorial extensions

If this is right

  • If prime-joint detection works as claimed, contrastive learning for skeleton actions no longer needs predefined moving body parts; the selection becomes per-sequence and unsupervised.
  • The reported gains on NTU RGB+D 120 imply that the learned representation transfers better across subjects and camera setups, which the paper's linear, fine-tuned, semi-supervised, and transfer-learning results support.
  • Reconstruction-based self-supervision can be improved by masking density-selected joints rather than motion-selected or randomly selected joints, as STJD-MP's new state of the art on the NTU 120 X-set protocol indicates.
  • Because the STJD module is used only during pre-training, the improved encoder adds no inference-time computational cost over the baseline backbone.

Reading between the lines

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

  • The same density-selection mechanism could be applied to other structured sequence modalities, such as hand skeletons, animal poses, or human-object interaction graphs, where static context joints are known to matter.
  • The fixed threshold $\beta$ is the most fragile component; an adaptive per-sample or per-action threshold derived from the distribution of $\Delta D_t(r)$ could reduce the failures the paper reports for subtle hand actions like typing or playing with a phone.
  • The STJD module could be plugged into other self-supervised frameworks as a generic saliency mask, beyond the STGCN contrastive and MAMP reconstruction backbones tested here.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a spatio-temporal joint density (STJD) measure, computed in the embedding space of a momentum encoder, to detect "prime joints" that include both moving and static discriminative joints. These prime joints are used in two self-supervised frameworks: STJD-CL, a contrastive method that aligns full-sequence and prime-joint representations and contrasts prime with non-prime joints, and STJD-MP, which replaces MAMP's motion-based masking with a density-change-based masking strategy. The method is evaluated with linear, supervised fine-tuning, semi-supervised, and transfer-learning protocols on NTU RGB+D 60, NTU RGB+D 120, and PKUMMD, reporting consistent gains over ActCLR and MAMP, with a headline improvement of 3.5 and 3.6 percentage points over ActCLR on NTU RGB+D 120 X-sub and X-set, respectively.

Significance. If the reported results hold after the table inconsistency is resolved, the paper makes a useful contribution: it offers a learnable, unsupervised criterion for selecting discriminative joints beyond predefined motion-based body parts, and it demonstrates consistent gains over ActCLR and MAMP across multiple datasets and downstream tasks. The paired t-test in Table IX, the prime-joints versus prime-parts comparison in Table VII, and the qualitative visualizations in Figures 3 and 4 are appropriate supporting analyses. However, the significance is currently tempered by an unresolved conflict between the two principal results tables, the heuristic nature of the Taylor-expansion motivation, and the lack of sensitivity and stability analysis for the threshold beta and the self-referential detection loop. These issues must be addressed before the claimed 3.5 and 3.6 percentage-point improvements can be taken at face value.

major comments (4)
  1. [§IV, Tables I and II] The two principal results tables conflict for the headline result. Table I reports 3s-STJD-CL at 77.8% on NTU RGB+D 120 X-sub, and the text in §IV states that STJD-CL outperformed ActCLR by 3.5 percentage points (the ActCLR value is 74.3% in both tables). Table II reports the same three-stream linear-evaluation configuration at 77.1%, which changes the improvement to 2.8 percentage points. Because the abstract's "3.5 percentage points" is derived from the Table I value, the central quantitative claim is not reproducible from the manuscript as written. Please reconcile the two tables, report per-run results and standard deviations for at least the three-stream configuration, and provide a working code link so the correct value can be verified.
  2. [§III-A, Eqs. (3)–(5)] The Taylor-expansion motivation for ΔD_t(r) is not mathematically sound as written. In Eq. (3), each Gaussian term G(r;v_{t,i},h_i) is expanded around a joint v_{t,k} without a clear definition of k relative to the summation index i, and the derivative should be with respect to the expansion point, not a separately introduced joint. The step from Eq. (4) to Eq. (5) is also unjustified: the term ((r−v_{t+δt,k})−(r−v_{t,k})) simplifies to v_{t,k}−v_{t+δt,k}, i.e., M(v_k), so the appearance of M(v_k)+M(r) in Eq. (5) is algebraically unexplained. In addition, Algorithm 1 uses the absolute value |ΔD_t(r)| while Eq. (2) defines ΔD_t(r) without an absolute value. Since this derivation is the stated motivation for prime-joint detection, please either correct the derivation or explicitly present it as heuristic intuition that is validated by the ablation evidence in Table VII.
  3. [§III-C and §IV-B] The prime-joint detector is applied to the embedding produced by the momentum encoder, and the training objective L_CL + L_RCL then aligns the full-sequence representation with the prime-joint representation and contrasts prime with non-prime joints from that same embedding. This creates a self-referential training signal: the encoder is trained to make the joints selected by the detector (computed from that same encoder) more distinctive. The external benchmark results show that the framework works, but they do not establish that STJD identifies a stable, action-discriminative property rather than a property that is partly an artifact of the encoder's current state. Please provide a stability analysis—for example, the overlap of detected prime joints across different encoder initializations or training checkpoints—and report the sensitivity of final accuracy to the threshold β on NTU RGB+D 120 and PKUMMD, not only on NTU 60 X-view.
  4. [§IV-B] The hyperparameter β is chosen empirically on NTU RGB+D 60 X-view (β=0.65) and then fixed for all datasets without sensitivity analysis elsewhere. Because β directly determines the set of prime joints, and because the authors report that accuracy varies by roughly 0.5 percentage points over the range 0.5≤β≤0.8 on that single benchmark, the manuscript should demonstrate that the fixed threshold is not accidentally favorable on the other benchmarks. A compact table or figure showing linear-evaluation accuracy versus β for NTU RGB+D 120 and PKUMMD would address this concern and would also help readers understand the robustness of the prime-joint detector.
minor comments (6)
  1. [Abstract and Section I] The code link is given as "STJD" rather than an actual repository address; please provide the full URL or a stable identifier.
  2. [§III-A, Eq. (1)] The notation is inconsistent: v_{t,i} is treated as a vector while r_{t,j} is a scalar channel, and the product over j sums channels inside the Gaussian. Please clarify the indexing, for example by writing v_{t,i,j} for the j-th channel of joint v_{t,i}.
  3. [§III-B, Algorithm 1] Step 9 computes the absolute temporal change while Eq. (2) does not; please align the definition of ΔD_t(r) between the equation, the algorithm, and the subsequent text.
  4. [§IV-B] The sentence reporting "87.67±0.0017%" gives an implausibly small standard deviation; please clarify the number of runs and the source of the variance, and state whether this is the standard deviation over classifier initializations or over pretraining seeds.
  5. [Table V] There is a formatting error in the MacDiff row: "65.677.3" should be "65.6 77.3" for the 1% labels X-sub column; please correct this.
  6. [Various] There are several typos that should be corrected: "ActionLet" in Figure 1, "J primt" in Section III-C, and "AimtCLR" in Table IV and Table VI.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the claimed accuracy gains are measured on external benchmarks with the STJD module removed at inference, and the self-referential prime-joint bootstrap is a pretext, not a fitted prediction.

full rationale

The proposed STJD is a kernel-density measure computed on the momentum encoder's embedding (Eqs. 1-2, Algorithm 1); the resulting prime-joint mask steers the contrastive and reconstruction objectives (Sec. III-C, III-D). This is self-referential in the sense that the mask is produced by the same network family being trained, but it is not circular in the load-bearing sense: the headline claim is the frozen-encoder accuracy on NTU60, NTU120 and PKUMMD (Tables I-VI), obtained after removing the STJD module, so the reported numbers are not constructed to equal the STJD output. The threshold beta=0.65 is a hyperparameter selected on NTU60 X-view with a quoted stability range (87.67 +/- 0.0017% for 0.5 <= beta <= 0.8) and then fixed for all datasets; it is not fitted to the NTU120 X-sub/X-set labels that produce the 3.5/3.6 pp claim. No self-citation is load-bearing: references [4], [6], [9], and [44] are related-work or baseline citations, and the derivation rests on the paper's own Eqs. (1)-(7) plus the publicly adopted MoCoV2/ActCLR/MAMP frameworks. Two non-circular weaknesses should be noted separately: Eq. (5) contains an algebraic slip in the Taylor expansion, and Table I (77.8%) conflicts with Table II (77.1%) for the same 3s-STJD-CL NTU120 X-sub configuration, weakening the abstract's 3.5 pp statement; neither issue makes the derivation equivalent to its inputs. The paper's own limitation statement in Section V acknowledges accuracy drops for some actions, but this is an efficacy limitation, not circularity.

Assumptions & free parameters 3 free parameters · 3 assumptions · 2 invented entities

The method introduces the STJD metric and the prime joint concept as new constructs. It relies on a fixed threshold beta chosen on one benchmark, learnable bandwidths, and a heuristic Taylor expansion. The paper's core assumptions are domain-specific heuristics rather than proven facts.

free parameters (3)
  • Threshold beta = 0.65
    Chosen empirically on NTU RGB+D 60 X-view, where accuracy plateaus for beta in [0.5,0.8], then fixed for all datasets.
  • Learnable bandwidths h_i
    Per-joint bandwidth parameters in the kernel density function, learned during pre-training.
  • Momentum coefficient alpha and temperature tau
    Standard hyperparameters in MoCoV2-based contrastive learning; not central to the novelty.
assumptions (3)
  • ad hoc to paper Taylor expansion approximation in Eq. (5) is valid
    The expansion assumes smoothness and ignores higher-order terms; it is used to justify that density change captures motion and interaction.
  • domain assumption High temporal density change indicates discriminative joints
    The paper assumes that joints with large |ΔD_t| are action-related; the threshold beta implements this without theoretical proof.
  • domain assumption Gaussian kernel with learnable bandwidth captures joint interactions
    The choice of kernel and bandwidth is heuristic; other kernels are acknowledged but not tested.
invented entities (2)
  • Prime joints
    purpose: A subset of joints, both moving and static, that guide self-supervised learning.
    The concept is defined and used to construct the training losses; its discriminative power is only validated through downstream accuracy on the same benchmarks, not through independent measurements.
  • Spatio-temporal joint density (STJD)
    purpose: A measure of joint interactions over time, used to detect prime joints.
    A new metric proposed in this paper; no independent validation outside the proposed methods.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatio-Temporal Joint Density Driven Learning for Skeleton-Based Action Recognition." pith.science (2026). https://pith.science/paper/DZIS5KI4

@misc{pith2026250523012,
  author       = {Pith},
  title        = {Pith review of: Spatio-Temporal Joint Density Driven Learning for Skeleton-Based Action Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DZIS5KI4}},
  note         = {Machine review of arXiv:2505.23012}
}
read the original abstract

Traditional approaches in unsupervised or self supervised learning for skeleton-based action classification have concentrated predominantly on the dynamic aspects of skeletal sequences. Yet, the intricate interaction between the moving and static elements of the skeleton presents a rarely tapped discriminative potential for action classification. This paper introduces a novel measurement, referred to as spatial-temporal joint density (STJD), to quantify such interaction. Tracking the evolution of this density throughout an action can effectively identify a subset of discriminative moving and/or static joints termed "prime joints" to steer self-supervised learning. A new contrastive learning strategy named STJD-CL is proposed to align the representation of a skeleton sequence with that of its prime joints while simultaneously contrasting the representations of prime and nonprime joints. In addition, a method called STJD-MP is developed by integrating it with a reconstruction-based framework for more effective learning. Experimental evaluations on the NTU RGB+D 60, NTU RGB+D 120, and PKUMMD datasets in various downstream tasks demonstrate that the proposed STJD-CL and STJD-MP improved performance, particularly by 3.5 and 3.6 percentage points over the state-of-the-art contrastive methods on the NTU RGB+D 120 dataset using X-sub and X-set evaluations, respectively.

Figures

Figures reproduced from arXiv: 2505.23012 by the authors.

Figure 1
Figure 1. Comparison of actionlets and prime joints for action Drink from NTU RGB+D 60. An actionlet is selected on the basis of pre-defined parts using motion. In contrast, the prime joints are detected based on the evolution of the proposed STJD. The discriminative static head joints are failed to select in actionlet, but were successfully included in the prime joints. introduced extreme augmentations, with the aim of learn… view at source ↗
Figure 2
Figure 2. The network architecture of the proposed STJD-CL.Prime joints are detected via STJD. A two-stream network is used for contrastive learning, and the online stream is updated with gradients while the offline stream is updated via momentum. Adaptive transformation T1 is adopted from [19]. InfoNCE loss, LCL, is defined to contrast the representation of the entire skeleton Xq with that of the prime joints in Xk and LRCL … view at source ↗
Figure 3
Figure 3. Visualization of actionlet and prime joints: The green joints represent actionlet or prime joints in the respective sequence, and the purple joints belong to non-actionlet or non￾prime joints. The 0 highlights the irrelevant joints included in Actionlet. using prime joints, prime parts that are covered by any of the prime joints and actionlets. Note that five body parts: left hand, right hand, left leg, right leg, a… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: The t-SNE visualization of embeddings on the NTU RGB+D 60 X-view benchmark. The same randomly selected [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 46 canonical work pages

  1. [1]

    Rmpe: Regional multi-person pose estimation

    Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu Lu. Rmpe: Regional multi-person pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2334–2343, 2017

    work page 2017

  2. [2]

    Channel-wise topology refinement graph convolution for skeleton-based action recognition

    Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu. Channel-wise topology refinement graph convolution for skeleton-based action recognition. In2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 13339–13348, 2021

    work page 2021

  3. [3]

    Rgb-d-based human motion recog- nition with deep learning: A survey, 2018

    Pichao Wang, Wanqing Li, Philip Ogunbona, Jun Wan, and Sergio Escalera. Rgb-d-based human motion recog- nition with deep learning: A survey, 2018

    work page 2018

  4. [4]

    Ogunbona

    Shanaka Ramesh Gunasekara, Wanqing Li, Jack Yang, and Philip. Ogunbona. Joint temporal pooling for improving skeleton-based action recognition. In2023 International Conference on Digital Image Computing: Techniques and Applications (DICTA), 2023

    work page 2023

  5. [5]

    Motion saliency based hierarchical attention network for action recognition.Multimedia Tools and Applications, 82(3):4533–4550, 2023

    Zihui Guo, Yonghong Hou, Renyi Xiao, Chuankun Li, and Wanqing Li. Motion saliency based hierarchical attention network for action recognition.Multimedia Tools and Applications, 82(3):4533–4550, 2023

    work page 2023

  6. [6]

    Asynchronous joint-based tempo- ral pooling for skeleton-based action recognition.IEEE Transactions on Circuits and Systems for Video Technol- ogy, pages 1–1, 2024

    Shanaka Ramesh Gunasekara, Wanqing Li, Jack Yang, and Philip Ogunbona. Asynchronous joint-based tempo- ral pooling for skeleton-based action recognition.IEEE Transactions on Circuits and Systems for Video Technol- ogy, pages 1–1, 2024

    work page 2024

  7. [7]

    Fake it till you recognize it: Quality assessment for human action generative models.IEEE Transactions on Biometrics, Behavior, and Identity Science, 6(2):261–271, 2024

    Bruno Degardin, Vasco Lopes, and Hugo Proenc ¸a. Fake it till you recognize it: Quality assessment for human action generative models.IEEE Transactions on Biometrics, Behavior, and Identity Science, 6(2):261–271, 2024

    work page 2024

  8. [8]

    Contrastive learning from extremely augmented skeleton sequences for self- supervised action recognition

    Tianyu Guo, Hong Liu, Zhan Chen, Mengyuan Liu, Tao Wang, and Runwei Ding. Contrastive learning from extremely augmented skeleton sequences for self- supervised action recognition. InProceedings of the AAAI conference on artificial intelligence, volume 36, pages 762–770, 2022

    work page 2022

Show all 52 references
  1. [9]

    Chuankun Li, Shuai Li, Yanbo Gao, Xingyu Gao, Ping Chen, Jian Li, and Wanqing Li. Unsupervised feature enrichment and fidelity preservation learning framework for skeleton-based action recognition.IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1, 2025

    2025

  2. [10]

    Skeletonmae: Graph-based masked autoencoder for skeleton sequence pre-training

    Hong Yan, Yang Liu, Yushen Wei, Zhen Li, Guanbin Li, and Liang Lin. Skeletonmae: Graph-based masked autoencoder for skeleton sequence pre-training. InPro- ceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5606–5618, October 2023

    2023

  3. [11]

    Masked motion pre- dictors are strong 3d action representation learners

    Yunyao Mao, Jiajun Deng, Wengang Zhou, Yao Fang, Wanli Ouyang, and Houqiang Li. Masked motion pre- dictors are strong 3d action representation learners. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10147–10157. IEEE, 2023

    2023

  4. [12]

    Ms2l: Multi-task self-supervised learning for skeleton based action recognition

    Lilang Lin, Sijie Song, Wenhan Yang, and Jiaying Liu. Ms2l: Multi-task self-supervised learning for skeleton based action recognition. InProceedings of the 28th ACM International Conference on Multimedia, MM ’20, page 2490–2498, New York, NY , USA, 2020. Associa- tion for Comp...

    2020

  5. [13]

    Motion-transformer: self-supervised pre-training for skeleton-based action recognition

    Yi-Bin Cheng, Xipeng Chen, Dongyu Zhang, and Liang Lin. Motion-transformer: self-supervised pre-training for skeleton-based action recognition. InProceedings of the 2nd ACM International Conference on Multimedia in Asia, MMAsia ’20, New York, NY , USA, 2021. Association for Co...

    2021

  6. [14]

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

    Mohamed Abdelfattah and Alexandre Alahi. S-jepa: A joint embedding predictive architecture for skeletal action recognition. In Ale ˇs Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G ¨ul Varol, editors,Computer Vision – ECCV 2024, pages 367–384, Ch...

    2024

  7. [15]

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

    Wenhan Wu, Yilei Hua, Ce Zheng, Shiqian Wu, Chen Chen, and Aidong Lu. Skeletonmae: Spatial-temporal masked autoencoders for self-supervised skeleton action recognition. In2023 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 224– 229, 2023

    2023

  8. [16]

    Contrastive learning from spatio-temporal mixed skeleton sequences for self- supervised skeleton-based action recognition

    Chen Zhan, Liu Hong, Guo Tianyu, Chen Zhengyan, Song Pinhao, and Tang Hao. Contrastive learning from spatio-temporal mixed skeleton sequences for self- supervised skeleton-based action recognition. InarXiv, 2022

    2022

  9. [17]

    Augmented skeleton based contrastive action learning with momentum lstm for unsupervised action recognition.Information Sciences, 569:90–109, 2021

    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

    2021

  10. [18]

    Bulat Khaertdinov, Stylianos Asteriadis, and Esam Ghaleb. Dynamic temperature scaling in contrastive self-supervised learning for sensor-based human activity recognition.IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(4):498–507, 2022

    2022

  11. [19]

    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 . In2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2363–2372, Los Alamitos, CA, USA, June 2023. IEEE Comp...

    2023

  12. [20]

    Sequence of the most in- formative joints (smij): A new representation for human skeletal action recognition

    Ferda Ofli, Rizwan Chaudhry, Gregorij Kurillo, Ren ´e Vidal, and Ruzena Bajcsy. Sequence of the most in- formative joints (smij): A new representation for human skeletal action recognition. In2012 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Work...

    2012

  13. [21]

    IEEE TRANSACTIONS ON BIOMETRICS, BEHA VIOR, AND IDENTITY SCIENCE 11 Mining actionlet ensemble for action recognition with depth cameras

    Jiang Wang, Zicheng Liu, Ying Wu, and Junsong Yuan. IEEE TRANSACTIONS ON BIOMETRICS, BEHA VIOR, AND IDENTITY SCIENCE 11 Mining actionlet ensemble for action recognition with depth cameras. In2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 1290–1297, 2012

    2012

  14. [22]

    Unsupervised visual representation learning by context prediction

    Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. InProceedings of the IEEE international conference on computer vision, pages 1422–1430, 2015

    2015

  15. [23]

    Unsupervised learning of visual representations by solving jigsaw puzzles

    Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69–84. Springer, 2016

    2016

  16. [24]

    Color- ful image colorization

    Richard Zhang, Phillip Isola, and Alexei A Efros. Color- ful image colorization. InComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Nether- lands, October 11-14, 2016, Proceedings, Part III 14, pages 649–666. Springer, 2016

    2016

  17. [25]

    Unsupervised representation learning by predicting im- age rotations.arXiv preprint arXiv:1803.07728, 2018

    Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting im- age rotations.arXiv preprint arXiv:1803.07728, 2018

    2018 arXiv

  18. [26]

    S4l: Self-supervised semi-supervised learn- ing

    Xiaohua Zhai, Avital Oliver, Alexander Kolesnikov, and Lucas Beyer. S4l: Self-supervised semi-supervised learn- ing. InProceedings of the IEEE/CVF international conference on computer vision, pages 1476–1485, 2019

    2019

  19. [27]

    Unsupervised representation learning by sorting sequences

    Hsin-Ying Lee, Jia-Bin Huang, Maneesh Singh, and Ming-Hsuan Yang. Unsupervised representation learning by sorting sequences. InProceedings of the IEEE international conference on computer vision, pages 667– 676, 2017

    2017

  20. [28]

    Discriminative hierarchical rank pooling for activity recognition

    Basura Fernando, Peter Anderson, Marcus Hutter, and Stephen Gould. Discriminative hierarchical rank pooling for activity recognition. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1924–1932, 2016

    1924

  21. [29]

    Self-supervised spatio- temporal representation learning for videos by predicting motion and appearance statistics

    Jiangliu Wang, Jianbo Jiao, Linchao Bao, Shengfeng He, Yunhui Liu, and Wei Liu. Self-supervised spatio- temporal representation learning for videos by predicting motion and appearance statistics. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogniti...

    2019

  22. [30]

    Improved baselines with momentum contrastive learning

    Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020

    2003 arXiv

  23. [31]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. InInternational con- ference on machine learning, pages 1597–1607. PMLR, 2020

    2020

  24. [32]

    Bootstrap your own latent-a new approach to self-supervised learn- ing.Advances in neural information processing systems, 33:21271–21284, 2020

    Jean-Bastien Grill, Florian Strub, Florent Altch ´e, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learn- ing.Advances in neura...

    2020

  25. [33]

    Exploring simple siamese representation learning

    Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750–15758, 2021

    2021

  26. [34]

    Barlow twins: Self-supervised learning via redundancy reduction

    Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and St´ephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. InInternational Conference on Machine Learning, pages 12310–12320. PMLR, 2021

    2021

  27. [35]

    Self-supervised action representation learn- ing from partial spatio-temporal skeleton sequences

    Yujie Zhou, Haodong Duan, Anyi Rao, Bing Su, and Jiaqi Wang. Self-supervised action representation learn- ing from partial spatio-temporal skeleton sequences. In Proceedings of the Thirty-Seventh AAAI Conference on Artificial Intelligence and Thirty-Fifth Conference on In- nov...

    2023

  28. [36]

    Spatial temporal graph convolutional networks for skeleton- based action recognition

    Sijie Yan, Yuanjun Xiong, and Dahua Lin. Spatial temporal graph convolutional networks for skeleton- based action recognition. InProceedings of the Thirty-Second AAAI Conference on Artificial Intelli- gence and Thirtieth Innovative Applications of Artifi- cial Intelligence Con...

    2018

  29. [37]

    Unsupervised repre- sentation learning with long-term dynamics for skeleton based action recognition.Proceedings of the AAAI Conference on Artificial Intelligence, 32(1), Apr

    Nenggan Zheng, Jun Wen, Risheng Liu, Liangqu Long, Jianhua Dai, and Zhefeng Gong. Unsupervised repre- sentation learning with long-term dynamics for skeleton based action recognition.Proceedings of the AAAI Conference on Artificial Intelligence, 32(1), Apr. 2018

    2018

  30. [38]

    Predict & cluster: Unsupervised skeleton based action recognition

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

    2020

  31. [39]

    Siyuan Yang, Jun Liu, Shijian Lu, Meng Hwa Er, and Alex C. Kot. Skeleton cloud colorization for unsuper- vised 3d action representation learning. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13423–13433, October 2021

    2021

  32. [40]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll´ar, and Ross Girshick. Masked autoencoders are scalable vision learners. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16000–16009, June 2022

    2022

  33. [41]

    Macdiff: Unified skeleton modeling with masked conditional diffusion

    Lehong Wu, Lilang Lin, Jiahang Zhang, Yiyang Ma, and Jiaying Liu. Macdiff: Unified skeleton modeling with masked conditional diffusion. In Ale ˇs Leonardis, Elisa Ricci, Stefan Roth, Olga Russakovsky, Torsten Sattler, and G ¨ul Varol, editors,Computer Vision – ECCV 2024, pages...

    2024

  34. [42]

    Idempotent unsupervised representation learning for skeleton-based action recognition

    Lilang Lin, Lehong Wu, Jiahang Zhang, and Jiaying Liu. Idempotent unsupervised representation learning for skeleton-based action recognition. InComputer Vision – ECCV 2024: 18th European Conference, Milan, Italy, September 29–October 4, 2024, Proceedings, Part XXVI, page 75–92...

    2024

  35. [43]

    Skeleton-contrastive 3d action representation learning

    Fida Mohammad Thoker, Hazel Doughty, and Cees GM Snoek. Skeleton-contrastive 3d action representation learning. InProceedings of the 29th ACM international conference on multimedia, pages 1655–1663, 2021. IEEE TRANSACTIONS ON BIOMETRICS, BEHA VIOR, AND IDENTITY SCIENCE 12

    2021

  36. [44]

    Contrastive positive mining for unsuper- vised 3d action representation learning

    Haoyuan Zhang, Yonghong Hou, Wenjing Zhang, and Wanqing Li. Contrastive positive mining for unsuper- vised 3d action representation learning. In Shai Avi- dan, Gabriel Brostow, Moustapha Ciss ´e, Giovanni Maria Farinella, and Tal Hassner, editors,Computer Vision – ECCV 2022, p...

    2022

  37. [45]

    3d human action representation learning via cross-view consistency pur- suit

    Linguo Li, Minsi Wang, Bingbing Ni, Hang Wang, Jiancheng Yang, and Wenjun Zhang. 3d human action representation learning via cross-view consistency pur- suit. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4741– 4750, 2021

    2021

  38. [46]

    Ntu rgb+ d: A large scale dataset for 3d human activity analysis

    Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang Wang. Ntu rgb+ d: A large scale dataset for 3d human activity analysis. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1010– 1019, 2016

    2016

  39. [47]

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

    Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, Alex Kot, • Liu, and L.-Y Duan. Ntu rgb+d 120: A large-scale benchmark for 3d human ac- tivity understanding. InIEEE Transactions on Pattern Analysis and Machine Intelligence. Institute of Electrical and Electro...

    2020

  40. [48]

    Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding

    Chunhui Liu, Yueyu Hu, Yanghao Li, Sijie Song, and Jiaying Liu. Pku-mmd: A large scale benchmark for continuous multi-modal human action understanding. In arXiv preprint arXiv:1703.07475, 2017

    2017 arXiv

  41. [49]

    Prototypical contrast and reverse prediction: Unsupervised skeleton based action recognition.IEEE Transactions on Multimedia, 25:624–634, 2023

    Shihao Xu, Haocong Rao, Xiping Hu, Jun Cheng, and Bin Hu. Prototypical contrast and reverse prediction: Unsupervised skeleton based action recognition.IEEE Transactions on Multimedia, 25:624–634, 2023

    2023

  42. [50]

    Unsupervised 3d human pose representation with viewpoint and pose disentanglement

    Qiang Nie, Ziwei Liu, and Yunhui Liu. Unsupervised 3d human pose representation with viewpoint and pose disentanglement. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part XIX 16, pages 102–118. Springer, 2020

    2020

  43. [51]

    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. InCom- puter Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part IV, page 209–225, Be...

    2022

  44. [52]

    Statistical comparisons of classifiers over multiple data sets.J

    Janez Dem ˇsar. Statistical comparisons of classifiers over multiple data sets.J. Mach. Learn. Res., 7:1–30, December 2006. Shanaka Ramesh Gunasekara(Member, IEEE) received the B.Sc. (hons) degree in electrical and electronic engineering from the University of Per- adeniya, Sr...

    2006

Pith tools

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