Pith. sign in

REVIEW 5 major objections 6 minor 49 references

Learning Adaptive Node Selection with External Attention for Human Interaction Recognition

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that skeleton-based interaction recognition is best served by first selecting the joints most active in the interaction and then applying cross-person external attention only to those joints, reporting top accuracy on…

desk verdict The node-selection idea is new and the experiments are solid, but the 'learnable threshold' is actually fixed at 0.5 because the hard mask blocks gradient flow. read the letter →

arxiv 2507.03936 v2 pith:DTOMKJVL submitted 2025-07-05 cs.CV

Chen Pang , Xuequan Lu , Qianyu Zhou , Lei Lyu This is my paper
classification cs.CV
keywords humaninteractionrecognitionskeleton-basedactiongraphconvolutionalnetworkattentionmechanismadaptivenodeselectionexternaltwo-person
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 tries to establish that for recognizing interactions between two people from skeleton data, a network should first decide which joints actually matter for the interaction and then let each person attend to the other person's joints that passed that selection. The proposed ASEA network models each person with a GCN, scores every joint by a temporal-weighted L2 norm of its learned features, keeps only joints above a learned threshold, and applies an external attention module across the two people's selected joints. On three benchmarks (NTU-26, SBU, Kinetics-10) it reports accuracy above existing two-person GCN baselines, suggesting that selecting informative joints before cross-person attention is more effective than predefined interaction adjacency matrices. The significance would be a generic way to replace fixed inter-person graph assumptions with a data-driven selection that changes per action and per sample.

What carries the argument

The load-bearing mechanism is the combination of the AT-NAC node-selection module and the External Attention module. AT-NAC computes per-frame joint energy E_{b,t,n}=||X_{b,:,t,n}||_2, derives temporal weights W_{b,t} from the softmax of frame-level variance of that energy, and obtains node amplitude S_{b,n}=sum_t W_{b,t} E_{b,t,n}. A learnable parameter $\alpha$ sets a sample-specific threshold tau_b = mu_b + $\alpha$ sigma_b, and joints with S_{b,n} > tau_b are the active nodes. External Attention then uses shared query, key, and value projections to compute cross-person attention maps between active nodes of the two individuals, applying each attention map to the other person's values with a residual connection. The selection discards low-amplitude joints so that attention concentrates on interaction-critical regions, while the regularization loss L_reg = $\lambda$($\alpha$ - alpha_target)^2 keeps $\alpha$ from drifting.

What would settle it

A decisive test would be to run ASEA on an interaction benchmark where a human annotator marks the joints that actually drive each action (for example, the right hand of the person taking a photo, or the face of the person looking at the other) and compare the model's selected nodes against those marks. If the model frequently fails to select the annotated interaction-critical joints while accuracy stays high, or if an oracle variant that forces selection of the annotated joints does not beat AT-NAC, then the amplitude assumption is not the source of the reported gains.

Watch

Extended reading notes

Core claim

The paper's central claim is that interaction relationships between two skeletons can be captured more effectively and flexibly by selecting a small set of 'active' joints and applying external attention only to them. The authors argue that predefined interaction adjacency matrices impose rigid, action-independent connections, whereas their Adaptive Temporal Node Amplitude Calculation (AT-NAC) computes, for each joint, an amplitude equal to a softmax-weighted sum over frames of the joint's L2 feature energy, with frame weights derived from the across-joint variance of that energy. A learnable threshold tau = mu + alpha*sigma then selects joints above it. The External Attention module attends from every selected node of one person to all selected nodes of the other, in both directions, and merges results through a residual connection. With this pipeline ASEA reports 90.52% and 91.77% on NTU-26 cross-subject and cross-setup, 99.82% average accuracy on SBU, and 58.18% Top-1 on Kinetics-10, all above the listed baselines.

Load-bearing premise

The core premise is that the temporal-weighted L2 norm of a joint's learned features reliably ranks how much that joint contributes to the interaction, so joints below the threshold can be dropped without losing information the external attention would need.

Editorial extensions

If this is right

  • If the reported numbers hold, hand-designed inter-person adjacency matrices are not required for strong two-person interaction recognition; selecting active joints per sample and applying external attention to them suffices.
  • The AT-NAC and external-attention modules are portable: adding them to five different GCN baselines on Kinetics-10 improves Top-1 accuracy in every case.
  • Because selection is per sample and per action, the model should automatically shift focus between action types, such as wrists and elbows for handshaking and legs for photo-taking.
  • Restricting cross-person attention to the selected active joints reduces the amount of noisy or redundant interaction modeling, consistent with the reported lower parameter count and FLOPs relative to the second-best method on Kinetics-10.

Reading between the lines

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

  • Editorial inference: the selection rule could be tested directly as a differentiable hard mask instead of a threshold, to separate the benefit of the amplitude criterion from the benefit of simply pruning noise.
  • Editorial inference: the frame-level variance weight assumes that informative frames are those where joints differ from each other; a low-variance but semantically crucial frame could be suppressed, so a targeted test would add a fixed bias to the variance weight and check accuracy on gaze-heavy interaction classes.
  • Editorial inference: the same adaptive node selection may transfer to group activity recognition with more than two people, where per-person active joints could gate pairwise attention without an O(n^2) predefined graph.
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

5 major / 6 minor

Summary. The paper proposes ASEA, a graph-convolutional network for skeleton-based human interaction recognition. The model processes each individual with an intra-person GCN, estimates per-joint activity using a temporal-weighted L2 norm of learned features (AT-NAC), selects active joints via a sample-specific threshold with a supposedly trainable sensitivity parameter alpha, and then applies an external attention mechanism between the two individuals on the selected nodes. Experiments on a custom NTU-26 subset, SBU, and a custom Kinetics-10 subset report state-of-the-art or near-state-of-the-art accuracy, with ablations supporting the contribution of the AT-NAC and external attention modules.

Significance. If the described mechanism works as claimed, the paper offers a plausible alternative to predefined inter-person adjacency matrices: it selects interaction-critical joints dynamically and then models cross-person attention only on those joints. The code availability and the consistent ablations are strengths. However, the central 'learnable threshold' claim is compromised by the non-differentiable hard mask, and the small reported margins over strong baselines, together with the absence of error bars and ambiguity about the backbone configuration, weaken the SOTA claim. The problem is relevant, and the architecture is reasonable, but the technical gap must be addressed before the contribution can be accepted.

major comments (5)
  1. [Sec. 3.4, Eqs. (7)-(8)] The hard threshold mask M_{b,n} = 1[S_{b,n} > mu_b + alpha*sigma_b] is piecewise constant in alpha. For continuous feature distributions, an infinitesimal change in alpha changes the mask only on a measure-zero set, so the task-loss gradient dL_task/dalpha is zero almost everywhere. The only gradient on alpha comes from L_reg in Eq. (11), which pulls alpha back to its initialization 0.5. Consequently, alpha is effectively fixed at 0.5 unless the paper describes a straight-through estimator, a soft relaxation, or another gradient path; no such mechanism is mentioned in Sec. 3.4-3.6. This undermines the claimed 'learnable threshold' and the flexibility attributed to AT-NAC. Please either provide the missing differentiable selection mechanism or revise the contribution and the associated text to state that the threshold is fixed at mu_b + 0.5*sigma_b.
  2. [Sec. 4.1 and Tables 1-3] The paper introduces custom subsets (NTU-26 and Kinetics-10) and reports no error bars, standard deviations, or significance tests. The reported improvements over the nearest baselines are small (Table 1: 90.52 vs 90.20 on X-Sub; Table 2: 99.82 vs 99.10; Table 3: 58.18 vs 57.58 Top-1). Without multiple runs or variance information, the state-of-the-art claim is not statistically supported. Please report mean and standard deviation over at least three independent runs, and where feasible, re-run the compared baselines under the same subset and protocol.
  3. [Sec. 3.2 and Table 3] The relationship between the 'ASEA (Ours)' row and the 'Baseline + AEA' rows in Table 3 is unclear. The method section says the intra-GCN is based on channel-wise topology (CTR-GCN), yet CTR-GCN + AEA (56.57 Top-1) is well below ASEA (58.18 Top-1). The paper does not specify which backbone ASEA uses, how the multi-scale temporal modeling interacts with the AEA module, or what additional components distinguish ASEA from simply adding AEA to a known baseline. This ambiguity prevents the reader from attributing the observed gains to the proposed modules. Please clarify the exact configuration of ASEA used for each table.
  4. [Abstract and Sec. 3.3] The claim that ASEA captures interaction relationships 'without predefined assumptions' is stronger than what the method actually does. The intra-person GCN uses a fixed skeleton adjacency matrix A in R^{N x N} (Sec. 3.3), which is a predefined anatomical prior. The method avoids predefined inter-person interaction matrices, but it still relies on a skeleton prior. Please revise the wording to say 'without predefined inter-person interaction adjacency matrices' or an equivalent qualifier.
  5. [Sec. 4.3, Tables 4-5] The ablations show that AT-NAC adds 0.73% (X-Sub) and 0.27% (X-Set) over applying external attention to all nodes, and the comparison against node velocity (76.35% vs 90.52% on X-Sub) is an unusually large gap that suggests the velocity-based selection is not a competitive alternative. To support the claim that the learnable threshold matters, please include a sensitivity analysis with alpha fixed at several values (e.g., 0.2, 0.5, 0.8) and compare against the trained alpha result. This would also clarify whether the selection mechanism's benefit comes from the threshold value or from the sample-dependent statistics mu_b and sigma_b.
minor comments (6)
  1. [Sec. 3.5] The section heading uses 'Extra Attention' while the abstract and the rest of the paper consistently use 'External Attention'; please unify the terminology.
  2. [Fig. 2 and Sec. 3.1] The input tensor shape is written as X in R^{T x C x M x N} in Fig. 2 but defined as R^{C x T x M x N} in Sec. 3.1; please align the notation.
  3. [Sec. 4.1] The dataset name is spelled 'Kinetic-10' in the text and 'Kinetics-10' in the tables; please use one consistent spelling throughout.
  4. [References] References [8] and [9] point to the same paper (AIGCN); please merge or disambiguate them.
  5. [Sec. 4.3, Table 4] The caption says 'AT-NC' instead of 'AT-NAC'; please fix the typo.
  6. [Sec. 3.6] The text states that lambda is determined via cross-validation, but no value or sensitivity analysis is reported; please give the selected value and, ideally, a small study of its effect.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ASEA's reported accuracies are external benchmark results, and the paper's self-citations are not load-bearing.

full rationale

The paper makes no first-principles derivation whose output is equivalent to its input. Its central claims are test accuracies on NTU-26, SBU, and Kinetics-10, measured end-to-end against external baselines; these are not quantities fitted from the same data. The AT-NAC selection uses features produced by the same network that consumes the selected nodes, which is ordinary joint training rather than a definitional loop. The multi-scale temporal module is credited to the authors' prior work [24], and [23]/[25] are cited as prior GCN examples, but neither citation supplies the central novelty (adaptive node selection plus cross-person attention); thus the self-citations are not load-bearing. The skeptic's point that the hard mask in Eq. (8) blocks task-loss gradients to alpha in Eq. (7) is a concrete architectural and testability concern about whether alpha is genuinely learned, but it is not circular: it would show a claimed mechanism may be ineffective, not that any reported prediction reduces by construction to a fitted input. No equation is defined in terms of its own output, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The free parameters are the learnable threshold coefficient alpha, the softmax temperature gamma, and the regularization weight lambda. The major assumptions are the reliability of the feature-norm activity measure for node selection, the sufficiency of frame-level cross-attention between the selected joints, and the use of a predefined intra-person skeleton prior. No invented entities are introduced.

free parameters (3)
  • alpha (threshold sensitivity) = trained, initialized to 0.5, regularized to 0.5
    Learnable parameter in Eq. (7) that scales the standard deviation to form the node-selection threshold. Its value is fit during training, and the regularization loss in Eq. (11) keeps it near 0.5.
  • gamma (temporal softmax temperature) = unspecified
    Temperature in Eq. (5) controlling the sharpness of the temporal attention weights. The paper only states gamma > 0 and does not specify its value or whether it is learned, so it is a hand-chosen hyperparameter.
  • lambda (regularization weight) = determined via cross-validation
    Balancing coefficient in Eq. (11) that weights the regularization loss; chosen by cross-validation per the text in Section 3.6.
assumptions (4)
  • domain assumption The intra-person skeleton graph topology is predefined via natural human joint connections (from the GCN baseline), so the claim of 'no predefined assumptions' applies only to inter-person interactions.
    Section 3.3 uses a channel-wise topology GCN based on the human body structure, which is a predefined structural prior.
  • ad hoc to paper The temporal-weighted L2 norm of GCN node features is a reliable measure of a joint's involvement in the interaction, and joints below the learned threshold can be safely excluded.
    Eqs. (3)-(8) define this selection criterion; the paper provides a qualitative visualization (Fig. 3) but no quantitative evidence that discarded joints are truly non-informative.
  • domain assumption Cross-attention from each individual's active nodes to the other individual's active nodes, within the same frame, is sufficient to model interaction semantics.
    Eqs. (9)-(10) define the attention; this assumes frame-level correspondence and ignores longer-range temporal dependencies during attention.
  • ad hoc to paper The thresholds derived from per-sample mean and standard deviation of node amplitudes, with a safeguard to keep at least one node, produce a consistent selection across samples of the same action.
    Eq. (7)-(8) and the safeguard in Section 3.4; the paper argues for intra-class consistency but only shows two examples (Fig. 5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Adaptive Node Selection with External Attention for Human Interaction Recognition." pith.science (2026). https://pith.science/paper/DTOMKJVL

@misc{pith2026250703936,
  author       = {Pith},
  title        = {Pith review of: Learning Adaptive Node Selection with External Attention for Human Interaction Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DTOMKJVL}},
  note         = {Machine review of arXiv:2507.03936}
}
read the original abstract

Most GCN-based methods model interacting individuals as independent graphs, neglecting their inherent inter-dependencies. Although recent approaches utilize predefined interaction adjacency matrices to integrate participants, these matrices fail to adaptively capture the dynamic and context-specific joint interactions across different actions. In this paper, we propose the Active Node Selection with External Attention Network (ASEA), an innovative approach that dynamically captures interaction relationships without predefined assumptions. Our method models each participant individually using a GCN to capture intra-personal relationships, facilitating a detailed representation of their actions. To identify the most relevant nodes for interaction modeling, we introduce the Adaptive Temporal Node Amplitude Calculation (AT-NAC) module, which estimates global node activity by combining spatial motion magnitude with adaptive temporal weighting, thereby highlighting salient motion patterns while reducing irrelevant or redundant information. A learnable threshold, regularized to prevent extreme variations, is defined to selectively identify the most informative nodes for interaction modeling. To capture interactions, we design the External Attention (EA) module to operate on active nodes, effectively modeling the interaction dynamics and semantic relationships between individuals. Extensive evaluations show that our method captures interaction relationships more effectively and flexibly, achieving state-of-the-art performance.

Figures

Figures reproduced from arXiv: 2507.03936 by the authors.

Figure 1
Figure 1. Existing methods struggle with multi-person interactions and inefficient modeling. (a) Most GCN-based methods treat [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overview of our Active Node Selection with External Attention Network (ASEA). Given an input skeleton sequence, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Visualization of the selection of active nodes across [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of attention patterns for the “taking [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Visualization of node variations for different sam [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 47 canonical work pages

  1. [1]

    Farzaneh Askari, Cyril Yared, Rohit Ramaprasad, Devin Garg, Anjun Hu, and James J Clark. 2024. Video Interaction Recognition using an Attention Augmented Relational Network and Skeleton Data. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 3225–3234

    work page 2024

  2. [2]

    Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. 2017. Realtime multi- person 2d pose estimation using part affinity fields. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 7291–7299

    work page 2017

  3. [3]

    Joao Carreira and Andrew Zisserman. 2017. Quo vadis, action recognition? a new model and the kinetics dataset. Inproceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 6299–6308

    2017

  4. [4]

    Bohong Chen, Yumeng Li, Yao-Xiang Ding, Tianjia Shao, and Kun Zhou. 2024. Enabling synergistic full-body control in prompt-based co-speech motion gen- eration. InProceedings of the 32nd ACM International Conference on Multimedia. 6774–6783

    work page 2024

  5. [5]

    Yuxin Chen, Ziqi Zhang, Chunfeng Yuan, Bing Li, Ying Deng, and Weiming Hu

    work page

  6. [6]

    Hyung-gun Chi, Myoung Hoon Ha, Seunggeun Chi, Sang Wan Lee, Qixing Huang, and Karthik Ramani. 2022. Infogcn: Representation learning for human skeleton- based action recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20186–20196

    work page 2022

  7. [7]

    Yunfeng Diao, Baiqi Wu, Ruixuan Zhang, Ajian Liu, Xiaoshuai Hao, Xingxing Wei, Meng Wang, and He Wang. 2024. TASAR: Transfer-based Attack on Skeletal Action Recognition.arXiv preprint arXiv:2409.02483(2024)

    arXiv 2024

  8. [9]

    Feng Gao, Hailun Xia, and Zhihao Tang. 2022. Attention interactive graph convolutional network for skeleton-based human interaction recognition. In2022 IEEE International Conference on Multimedia and Expo (ICME). IEEE, 1–6

    work page 2022

Show all 49 references
  1. [10]

    Jungho Lee, Minhyeok Lee, Dogyoon Lee, and Sangyoun Lee. 2023. Hierarchically decomposed graph convolutional networks for skeleton-based action recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision. 10444– 10453

    2023

  2. [11]

    Chao Li, Qiaoyong Zhong, Di Xie, and Shiliang Pu. 2018. Co-occurrence feature learning from skeleton data for action recognition and detection with hierarchical aggregation. InProceedings of the 27th International Joint Conference on Artificial Intelligence. 786–792

    2018

  3. [12]

    Jianan Li, Xuemei Xie, Yuhan Cao, Qingzhe Pan, Zhifu Zhao, and Guangming Shi. 2021. Knowledge embedded GCN for skeleton-based two-person interaction recognition.Neurocomputing444 (2021), 338–348

    2021

  4. [13]

    Maosen Li, Siheng Chen, Xu Chen, Ya Zhang, Yanfeng Wang, and Qi Tian

  5. [14]

    Zhengcen Li, Yueran Li, Linlin Tang, Tong Zhang, and Jingyong Su. 2022. Two- person graph convolutional network for skeleton-based human interaction recog- nition.IEEE Transactions on Circuits and Systems for Video Technology33, 7 (2022), 3333–3342

    2022

  6. [15]

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

    2019

  7. [16]

    Jun Liu, Amir Shahroudy, Dong Xu, Alex C Kot, and Gang Wang. 2017. Skeleton- based action recognition using spatio-temporal LSTM network with trust gates. IEEE Transactions on Pattern Analysis and Machine Intelligence40, 12 (2017), 3007–3021

    2017

  8. [17]

    Mengyuan Liu, Chen Chen, Songtao Wu, Fanyang Meng, and Hong Liu. 2025. Learning mutual excitation for hand-to-hand and human-to-human interaction recognition.IEEE Transactions on Human-Machine Systems(2025)

    2025

  9. [18]

    Ziyu Liu, Hongwen Zhang, Zhenghao Chen, Zhiyong Wang, and Wanli Ouyang

  10. [19]

    Qianhui Men, Edmond SL Ho, Hubert PH Shum, and Howard Leung. 2021. A two- stream recurrent network for skeleton-based human interaction recognition. In 2020 25th International Conference on Pattern Recognition (ICPR). IEEE, 2771–2778

    2021

  11. [20]

    Woomin Myung, Nan Su, Jing-Hao Xue, and Guijin Wang. 2024. DeGCN: De- formable Graph Convolutional Networks for Skeleton-Based Action Recognition. IEEE Transactions on Image Processing33 (2024), 2477–2490

    2024

  12. [21]

    Woomin Myung, Nan Su, Jing-Hao Xue, and Guijin Wang. 2024. Degcn: De- formable graph convolutional networks for skeleton-based action recognition. IEEE Transactions on Image Processing33 (2024), 2477–2490

    2024

  13. [22]

    Xuan Son Nguyen. 2021. Geomnet: A neural network based on riemannian geometries of spd matrix space and cholesky space for 3d skeleton-based inter- action recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision. 13379–13389

    2021

  14. [23]

    Chen Pang, Xingyu Gao, Zhenyu Chen, and Lei Lyu. 2024. Self-Adaptive Graph With Nonlocal Attention Network for Skeleton-Based Action Recognition.IEEE Transactions on Neural Networks and Learning Systems35, 12 (2024), 17057–17069

    2024

  15. [24]

    Chen Pang, Xuequan Lu, and Lei Lyu. 2023. Skeleton-based action recognition through contrasting two-stream spatial-temporal networks.IEEE Transactions on Multimedia25 (2023), 8699–8711

    2023

  16. [25]

    Chen Pang, Xuequan Lu, and Lei Lyu. 2023. Skeleton-Based Action Recognition Through Contrasting Two-Stream Spatial-Temporal Networks.IEEE Transactions on Multimedia25 (2023), 8699–8711

    2023

  17. [26]

    Yunsheng Pang, Qiuhong Ke, Hossein Rahmani, James Bailey, and Jun Liu. 2022. Igformer: Interaction graph transformer for skeleton-based human interaction recognition. InEuropean Conference on Computer Vision. Springer, 605–622

    2022

  18. [27]

    Mauricio Perez, Jun Liu, and Alex C Kot. 2021. Interaction relational network for mutual action recognition.IEEE Transactions on Multimedia24 (2021), 366–376

    2021

  19. [28]

    Helei Qiu, Biao Hou, Bo Ren, and Xiaohua Zhang. 2023. Spatio-temporal segments attention for skeleton-based action recognition.Neurocomputing518 (2023), 30– 38

    2023

  20. [29]

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

    2019

  21. [30]

    Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. 2020. Decoupled spatial- temporal attention network for skeleton-based action-gesture recognition. In Proceedings of the Asian Conference on Computer Vision

    2020

  22. [31]

    Lei Shi, Yifan Zhang, Jian Cheng, and Hanqing Lu. 2020. Skeleton-based action recognition with multi-stream adaptive graph convolutional networks.IEEE Transactions on Image Processing29 (2020), 9532–9545

    2020

  23. [32]

    Haitao Tian and Pierre Payeur. 2024. Stitch Contrast and Segment_Learning a Human Action Segmentation Model Using Trimmed Skeleton Videos.arXiv preprint arXiv:2412.14988(2024)

    2024 arXiv

  24. [33]

    Guoquan Wang, Mengyuan Liu, Hong Liu, Peini Guo, Ti Wang, Jingwen Guo, and Ruijia Fan. 2024. Augmented skeleton sequences with hypergraph network for self-supervised group activity recognition.Pattern Recognition152 (2024), 110478

    2024

  25. [34]

    Lei Wang and Piotr Koniusz. 2023. 3mformer: Multi-order multi-mode trans- former for skeletal action recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5620–5631

    2023

  26. [35]

    Shengqin Wang, Yongji Zhang, Minghao Zhao, Hong Qi, Kai Wang, Fenglin Wei, and Yu Jiang. 2022. Skeleton-based action recognition via temporal-channel aggregation.arXiv preprint arXiv:2205.15936(2022)

    2022 arXiv

  27. [36]

    Yuhang Wen, Zixuan Tang, Yunsheng Pang, Beichen Ding, and Mengyuan Liu

  28. [37]

    Wangmeng Xiang, Chao Li, Yuxuan Zhou, Biao Wang, and Lei Zhang. 2022. Lan- guage supervised training for skeleton-based action recognition.arXiv preprint arXiv:2208.053187 (2022)

    2022 arXiv

  29. [38]

    Sijie Yan, Yuanjun Xiong, and Dahua Lin. 2018. Spatial temporal graph convolu- tional networks for skeleton-based action recognition. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 32

    2018

  30. [39]

    Chao-Lung Yang, Aji Setyoko, Hendrik Tampubolon, and Kai-Lung Hua. 2020. Pairwise adjacency matrix on spatial temporal graph convolution network for skeleton-based two-person interaction recognition. In2020 IEEE International Conference on Image Processing (ICIP). IEEE, 2166–2...

    2020

  31. [40]

    Kiwon Yun, Jean Honorio, Debaleena Chattopadhyay, Tamara L Berg, and Dim- itris Samaras. 2012. Two-person interaction detection using body-pose features and multiple instance learning. In2012 IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops...

    2012

  32. [41]

    Qinyang Zeng, Ronghao Dang, Xun Zhou, Chengju Liu, and Qijun Chen. 2025. Contrastive Feedback Vision-Language for 3D Skeleton-Based Action Recogni- tion.IEEE Transactions on Multimedia(2025)

    2025

  33. [42]

    Pengfei Zhang, Cuiling Lan, Junliang Xing, Wenjun Zeng, Jianru Xue, and Nan- ning Zheng. 2019. View adaptive neural networks for high performance skeleton- based human action recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence41, 8 (2019), 1963–1978

    2019

  34. [43]

    Zhengyou Zhang. 2012. Microsoft kinect sensor and its effect.IEEE multimedia 19, 2 (2012), 4–10

    2012

  35. [44]

    Liping Zhu, Bohua Wan, Chengyang Li, Gangyi Tian, Yi Hou, and Kun Yuan

  36. [45]

    Yan Zhuang, Yanlu Cai, Weizhong Zhang, and Cheng Jin. 2024. Future Motion Dynamic Modeling via Hybrid Supervision for Multi-Person Motion Prediction Uncertainty Reduction. InProceedings of the 32nd ACM International Conference on Multimedia. 9563–9572

    2024

  37. [49]

    Dyadic relational graph convolutional networks for skeleton-based human interaction recognition.Pattern Recognition115 (2021), 107920

    2021

  38. [2019]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Actional-structural graph convolutional networks for skeleton-based action recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 3595–3603

  39. [2020]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Disentangling and unifying graph convolutions for skeleton-based action recognition. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 143–152

  40. [2021]

    InProceedings of the IEEE/CVF International Conference on Computer Vision

    Channel-wise topology refinement graph convolution for skeleton-based action recognition. InProceedings of the IEEE/CVF International Conference on Computer Vision. 13359–13368

  41. [2023]

    In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Interactive spatiotemporal token attention network for skeleton-based general interactive action recognition. In2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 7886–7892

Pith tools

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