Pith. sign in

REVIEW 3 major objections 4 minor 70 references

The paper claims that zero-shot skeleton action recognition improves when text-side embeddings are dynamically refined at test time; a confidence-guided affine update of LLM-generated, part-level descriptions yields state-of-the-art ZSL and

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

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

T0 review · deepseek-v4-flash

2026-08-03 16:51 UTC pith:HY57CNOK

load-bearing objection Real extension of PURLS with test-time text refinement; just don't trust the SOTA numbers until the transductive comparison is fixed. the 3 major comments →

arxiv 2512.11941 v2 pith:HY57CNOK submitted 2025-12-12 cs.CV cs.AI

DynaPURLS: Dynamic Refinement of Part-Aware Representations for Skeleton-Based Zero-Shot Action Recognition

Jingmin Zhu , Anqi Zhu , James Bailey , Jun Liu , Hossein Rahmani , Mohammed Bennamoun , Farid Boussaid , Qiuhong Ke This is my paper
classification cs.CV cs.AI
keywords zero-shot learningskeleton-based action recognitiontest-time adaptationvision-language alignmentcross-modal attentionpseudo-label refinementclass-balanced memory bankmulti-granularity semantics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Zero-shot skeleton action recognition asks a model trained on seen actions to recognize unseen actions from textual descriptions alone. The paper's claim is that the persistent seen-to-unseen gap is largely a semantic-side problem: static, class-level text embeddings fail to track the visual specifics of skeleton motion. DynaPURLS tests this by generating multi-granularity action descriptions (global movement, four body parts, three temporal phases), aligning each to skeleton joints via adaptive cross-modal attention, and then—at inference time—refining the text embeddings with a lightweight affine transform guided by the model's own confident predictions, stored in a class-balanced memory bank. The reported outcome is a clear jump over prior methods, with dynamic refinement adding up to 30 points of accuracy over the static version and final ZSL accuracy reaching 88–89% on standard splits. If correct, this makes test-time semantic adaptation a practical way to close the zero-shot domain gap.

Core claim

The central discovery is that refining the semantic anchors—not the visual encoder—is enough to bridge the train-test domain shift in skeleton zero-shot recognition. After training, DynaPURLS freezes the visual backbone and text encoder, and only learns a per-class scaling-and-bias transform on the text embeddings at test time, minimizing cross-entropy over high-confidence predictions. Because the text features double as attention queries, updating them changes which skeleton joints are aggregated, so a small affine update propagates through the whole fused representation. A class-balanced memory bank (at most sixteen high-confidence samples per class) keeps the pseudo-label set diverse and

What carries the argument

The load-bearing mechanism is affine query refinement: at test time the static text embedding matrix F is transformed into F′ = N(S ⊙ F + ΔF), where S and ΔF are learnable scale and bias tensors; a gradient-guided cross-entropy loss over confident, class-balanced pseudo-labels updates only these parameters. Since the refined embeddings serve as queries in the cross-modal attention that aggregates skeleton joints, the transformation effectively re-partitions visual evidence toward unseen classes. The class-balanced memory bank stabilizes this by keeping at most K high-confidence samples per class, and an entropy-based gating rule handles the seen/unseen imbalance of GZSL.

Load-bearing premise

The refinement step assumes that high-confidence predictions on test samples are usually correct, so that minimizing cross-entropy over those pseudo-labels moves the text embeddings toward the true unseen distribution; the paper itself notes the model can be confidently wrong and the memory bank may accumulate systematically biased samples, leading to semantic drift.

What would settle it

Run the method on a test stream with manipulated order—send all samples of a single unseen class first, then the rest—and compare accuracy to a shuffled stream. If accuracy drops well below the static baseline or seen-class accuracy erodes, the online refinement is depending on favorable sample order rather than on the class-balanced memory bank alone. Sharper still: annotate the high-confidence pseudo-labels; if per-class precision at the operating threshold is low enough that cross-entropy updates reinforce wrong classes, the refinement will drift.

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

If this is right

  • If the gains are real, test-time semantic refinement becomes a cheap standard add-on for zero-shot skeleton recognition: only the small scale/bias parameters are updated, leaving heavy encoders frozen.
  • The ablations show refinement pays off most when the initial alignment is fine-grained and adaptive, implying the benefit of test-time adaptation compounds with representation quality.
  • The memory bank is essential to stability: without it, pseudo-label refinement is unreliable, whereas with it, unseen-class accuracy in GZSL rises by up to about 25 points while seen-class accuracy holds.
  • The part-level (body-part) semantics contribute more to refinement than temporal-phase semantics, suggesting that concrete spatial descriptions provide the most transferable signal.
  • The framework reports consistent gains across several skeleton encoders and language-model backbones, so the mechanism is not tied to one architecture.

Where Pith is reading between the lines

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

  • The same confidence-guided semantic refinement could transfer to other zero-shot modalities—RGB video, audio, point clouds—wherever static class-level text embeddings are the bottleneck; the mechanism only assumes a text-query attention module.
  • A stronger stress test than the paper runs would manipulate test order (e.g., all samples of one class first) to see whether the online memory-bank update is order-sensitive; if it is, open-world streams with non-i.i.d. arrival could degrade.
  • Because the paper admits the model can be confidently wrong, a natural extension is to measure precision of high-confidence pseudo-labels per class and add a rejection or ensemble step when precision is low.
  • The GZSL entropy threshold is tuned on a validation set; a calibration scheme that removes that dependency would make the method more deployable.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper proposes DynaPURLS, an extension of the authors' earlier PURLS framework for zero-shot skeleton-based action recognition (ZS-SAR). It generates multi-granularity semantic descriptions with GPT-3, aligns skeleton features to these descriptions via adaptive cross-modal attention, and adds a test-time adaptation module (Algorithm 1) that learns per-class affine transformations S and ΔF of the text embeddings by minimizing cross-entropy on confident pseudo-labels, stabilized by a class-balanced memory bank. Experiments on NTU RGB+D 60/120 and PKU-MMD report large gains over published methods under fixed and random splits, including 88.52% ZSL on NTU60 55/5 and 89.06% on NTU120 110/10, and the paper claims new state-of-the-art results in both ZSL and GZSL settings.

Significance. The core idea—adapting semantic embeddings online to the unlabeled test stream rather than keeping them static—is a plausible and potentially valuable way to address the seen-unseen domain gap in ZS-SAR. The paper includes ablations showing that a class-balanced memory bank stabilizes pseudo-label optimization and that the refinement gains are not merely an artifact of the initial alignment quality. The release of source code is a positive aspect. However, the headline comparison is between a transductive algorithm and inductive baselines: the method's only novel component, test-time refinement, uses the test distribution during inference, whereas the compared methods do not. As presented, the central 'new state-of-the-art' claim is therefore not supported by the reported experiments.

major comments (3)
  1. [Tables 3–4 vs. Algorithm 1 / §3.4.2] The SOTA claim rests on a transductive evaluation. Algorithm 1 updates S and ΔF from the unlabeled test stream using Eq. (15)–(16) before returning predictions, while the baselines in Tables 3–4 (Neuron, SCoPLe, PURLS, etc.) are evaluated inductively. Table 6 shows the static 'Adaptive' row is 79.23/40.99/71.95/52.01, and the full DynaPURLS numbers are 88.52/71.80/89.06/69.11; the entire SOTA-level margin comes from the refinement step. Thus Tables 3–4 do not support a claim of state-of-the-art over inductive methods. Please report the inductive-only variant (refinement disabled) in the main comparison tables, or compare against transductive versions of the baselines, and state the protocol explicitly.
  2. [§3.4.2, Eq. (15); §6] The refinement objective minimizes cross-entropy on pseudo-labels produced by the model itself, creating a self-confirmation loop. Section 6 explicitly concedes that 'the model can be confidently wrong' and that the memory bank 'may accumulate systematically biased samples, potentially leading to semantic drift.' The reported gains are therefore partly a measure of how well the model's own confidence correlates with correctness on these benchmarks. The paper should quantify robustness to pseudo-label noise, for example by reporting an upper bound using ground-truth labels during refinement, a lower bound using corrupted labels, and error bars across runs or splits. Without this, the large margins in Tables 3–4 are difficult to interpret.
  3. [Tables 3–4] No error bars or standard deviations are reported. Fixed splits may be deterministic, but Table 4 reports 'averaged results' across random splits without variance. Given the very large gaps (e.g., NTU120 ZSL 90.04 vs. SCoPLe 53.34), the reader cannot assess whether the differences are statistically meaningful. Please provide standard deviations over the three random splits or multiple training seeds.
minor comments (4)
  1. [Algorithm 1] The condition 'if |M| ≥ Bmin' uses a variable Bmin that is never defined in Section 3.4.2 or the implementation details. Please specify its value or remove the ambiguity.
  2. [§4.1.4 vs. §5.7, Fig. 7] Implementation details set the confidence threshold to τ=0.1, but Section 5.7 states that performance improves as τ increases to 0.2 and Figure 7(a) shows a peak around 0.2. Please clarify which τ was used for the main results and reconcile the text with the figure.
  3. [Table 10] The PURLS baseline row reports 52.01 for both NTU120 110/10 and 96/24, but Table 3 shows PURLS ZSL of 71.95 on 110/10. This appears to be a typo; please correct.
  4. [§1, Contribution 1] The claim of being 'the first work to use test-time adaptation for zero-shot skeleton-based action recognition' is strong; please temper it or provide a more precise literature check, since generic TTA methods applied to skeleton ZSL are discussed in Section 2.3.

Circularity Check

1 steps flagged

Test-time refinement fits to the test stream; the reported SOTA numbers are transductive rather than inductive comparisons.

specific steps
  1. fitted input called prediction [Section 3.4.2, Eq. (15), Algorithm 1; results in Tables 3-4 and Table 6]
    "Compute refinement loss: Ladapt = − 1/|B| ∑_{(x_j, ŷ_j)∈B} log p′_{j,ŷ_j}; Update parameters: S←S−β∇_S Ladapt; Update parameters: ΔF←ΔF−β∇_{ΔF} Ladapt."

    The final test-set predictions are made after S and ΔF have been fit to that same test set: Algorithm 1 minimizes cross-entropy on pseudo-labels derived from the model's own current predictions (Eq. 14-15). The reported ZSL/GZSL numbers in Tables 3-4 are therefore not out-of-sample; Table 6 shows the entire SOTA margin over the static model comes from this refinement (+9.29/+30.81/+17.11/+17.10). The pseudo-label objective tends to reinforce the model's own confident choices, and Section 6 concedes the model 'can be confidently wrong' and that the memory bank 'may accumulate systematically biased samples, potentially leading to semantic drift.' The 'prediction' is thus partly a self-confirming fit to the test distribution, and comparing it to inductive baselines (Neuron, SCoPLe, PURLS) is

full rationale

The training-time multi-granularity alignment (Sections 3.2-3.3) is independently derived and externally benchmarked, and the PURLS self-citation is a normal precursor, not a load-bearing circular argument. No uniqueness theorem or ansatz is smuggled via self-citation. The one genuine circularity-like step is the inference-time refinement loop: Algorithm 1 adapts the semantic embeddings S,ΔF to the test stream using the model's own high-confidence pseudo-labels, and the same test samples are then used for evaluation. Because Table 6 attributes the entire reported SOTA gain to this refinement, the headline comparisons in Tables 3-4 are transductive results measured against inductive baselines, not independent predictions. The paper is transparent about using test-time adaptation, but the comparison is not protocol-matched, and the self-confirmation risk is acknowledged in Section 6. This warrants a partial circularity score rather than a claim that the derivation is definitionally circular.

Axiom & Free-Parameter Ledger

7 free parameters · 4 axioms · 0 invented entities

The main load-bearing external inputs are the LLM-generated descriptions, the CLIP/Shift-GCN embedding spaces, and the reliability of confident pseudo-labels. The paper's own limitation section concedes the last one can fail.

free parameters (7)
  • Test-time affine scale S = C×(P+Z+1)×d tensor, initialized to 1, optimized on test stream
    Eq. 12; fitted at inference on test samples via pseudo-label cross-entropy.
  • Test-time affine bias ΔF = C×(P+Z+1)×d tensor, initialized to 0, optimized on test stream
    Eq. 12; fitted at inference on test samples via pseudo-label cross-entropy.
  • Confidence threshold τ = 0.1
    Section 4.1.4 and Fig. 7(a); chosen by hand/validation to filter pseudo-labels.
  • Memory bank capacity K = 16
    Section 4.1.4 and Fig. 7(b); chosen by hand/validation to balance diversity and noise.
  • Refinement learning rate β = 0.01 with cosine decay
    Section 4.1.4; chosen for test-time optimization stability.
  • Entropy threshold δ for GZSL gating = tuned on a validation set
    Section 3.5; used to separate seen from unseen classes, but no reported value.
  • Number of body parts P and temporal segments Z = P=4, Z=3
    Section 3.2.1; fixed by design, not learned.
axioms (4)
  • domain assumption LLM-generated descriptions (GPT-3) accurately decompose skeleton actions into transferable body-part and temporal-phase concepts.
    Section 3.2.1 and Tables 1-2; the semantic side of the method depends on this decomposition being informative and stable.
  • domain assumption CLIP text embeddings and Shift-GCN skeleton features lie in a comparable semantic space and can be aligned by cross-modal attention.
    Section 3.3; the whole contrastive alignment rests on this cross-modal compatibility.
  • domain assumption High-confidence predictions on unseen classes are more often correct than random, and their pseudo-labels are reliable enough to adapt text embeddings.
    Section 3.4.2; the memory bank and refinement loss depend on this premise. The paper concedes failure cases in Section 6.
  • domain assumption Entropy threshold δ tuned on a validation set transfers to the test-time seen/unseen separation.
    Section 3.5; no unseen validation labels are available in standard GZSL, yet δ is tuned on a validation set.

pith-pipeline@v1.3.0-alltime-deepseek · 25066 in / 12769 out tokens · 108434 ms · 2026-08-03T16:51:59.680233+00:00 · methodology

0 comments
read the original abstract

Zero-shot skeleton-based action recognition (ZS-SAR) is fundamentally constrained by prevailing approaches that rely on aligning skeleton features with static, class-level semantics. This coarse-grained alignment fails to bridge the domain shift between seen and unseen classes, thereby impeding the effective transfer of fine-grained visual knowledge. To address these limitations, we introduce \textbf{DynaPURLS}, a unified framework that establishes robust, multi-scale visual-semantic correspondences and dynamically refines them at inference time to enhance generalization. Our framework leverages a large language model to generate hierarchical textual descriptions that encompass both global movements and local body-part dynamics. Concurrently, an adaptive partitioning module produces fine-grained visual representations by semantically grouping skeleton joints. To fortify this fine-grained alignment against the train-test domain shift, DynaPURLS incorporates a dynamic refinement module. During inference, this module adapts textual features to the incoming visual stream via a lightweight learnable projection. This refinement process is stabilized by a confidence-aware, class-balanced memory bank, which mitigates error propagation from noisy pseudo-labels. Extensive experiments on three large-scale benchmark datasets, including NTU RGB+D 60/120 and PKU-MMD, demonstrate that DynaPURLS significantly outperforms prior art, setting new state-of-the-art records. The source code is made publicly available at https://github.com/Alchemist0754/DynaPURLS

Figures

Figures reproduced from arXiv: 2512.11941 by Anqi Zhu, Farid Boussaid, Hossein Rahmani, James Bailey, Jingmin Zhu, Jun Liu, Mohammed Bennamoun, Qiuhong Ke.

Figure 1
Figure 1. Figure 1: An example illustrating the limitations of global-only alignment. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: t-SNE visualization of feature distributions before and after test [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Training pipeline of DynaPURLS. The model learns a robust static alignment between visual features extracted by the skeleton encoder and multi-granularity semantic embeddings from the text encoder. We employ a adaptive partitioning module where static text embeddings serve as queries to flexibly aggregate visual information. The entire network is optimized using a symmetric contrastive loss to align the re… view at source ↗
Figure 4
Figure 4. Figure 4: Static spatial partitioning scheme for decomposing 25 body joints [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Test-time feature refinement pipeline of DynaPURLS. The framework dynamically refines semantic embeddings based on high-confidence test samples stored in a class-balanced memory bank, enabling robust adaptation to distribution shifts in unseen classes. The refined embeddings are used to update the cross-modal attention mechanism for improved visual-semantic alignment. 3.3.2 Adaptive Partitioning via Cross-… view at source ↗
Figure 7
Figure 7. Figure 7: Hyperparameter sensitivity analysis on NTU RGB+D dataset. (a) [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Class-wise performance changes from PURLS to DynaPURLS on NTU RGB+D 120 (96/24 split). Actions are sorted by performance change, [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Confusion matrices comparison for NTU RGB+D 55/5 split before [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: t-SNE visualization of feature distributions during progressive [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

70 extracted references · 3 canonical work pages

  1. [1]

    3d semantic parsing of large-scale indoor spaces,

    I. Armeni, O. Sener, A. R. Zamir, H. Jiang, I. Brilakis, M. Fischer, and S. Savarese, “3d semantic parsing of large-scale indoor spaces,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016

    2016

  2. [2]

    Softgroup for 3d instance segmentation on point clouds,

    T. Vu, K. Kim, T. M. Luu, X. T. Nguyen, and C. D. Yoo, “Softgroup for 3d instance segmentation on point clouds,” 2022

    2022

  3. [3]

    Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection,

    Y. Li, A. W. Yu, T. Meng, B. Caine, J. Ngiam, D. Peng, J. Shen, B. Wu, Y. Lu, D. Zhou, Q. V . Le, A. Yuille, and M. Tan, “Deepfusion: Lidar-camera deep fusion for multi-modal 3d object detection,” 2022

    2022

  4. [4]

    Center-based 3d object de- tection and tracking,

    T. Yin, X. Zhou, and P . Krahenbuhl, “Center-based 3d object de- tection and tracking,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021, pp. 11 784–11 793

    2021

  5. [5]

    Deep learning for video-text retrieval: a review,

    C. Zhu, Q. Jia, W. Chen, Y. Guo, and Y. Liu, “Deep learning for video-text retrieval: a review,” International Journal of Multimedia Information Retrieval, vol. 12, no. 1, p. 3, 2023

    2023

  6. [6]

    Monocular 3d scene understanding with explicit occlusion reasoning,

    C. Wojek, S. Walk, S. Roth, and B. Schiele, “Monocular 3d scene understanding with explicit occlusion reasoning,” in 2011 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 06 2011, pp. 1993–2000

    2011

  7. [7]

    Multi-modal auto-encoders as joint estimators for robotics scene understanding,

    C. Cadena, A. Dick, and I. Reid, “Multi-modal auto-encoders as joint estimators for robotics scene understanding,” in Robotics: Science and Systems XII, 06 2016

    2016

  8. [8]

    A closer look at spatiotemporal convolutions for action recognition,

    D. Tran, H. Wang, L. Torresani, J. Ray, Y. LeCun, and M. Paluri, “A closer look at spatiotemporal convolutions for action recognition,” in CVPR. Computer Vision Foundation / IEEE Computer Society, 2018, pp. 6450–6459. [Online]. Available: http://openaccess.thecvf.com/content_cvpr_ 2018/html/Tran_A_Closer_Look_CVPR_2018_paper.html

    2018

  9. [9]

    Spatio-temporal pyramid graph convolutions for human action recognition and postural assessment,

    B. Parsa, A. Narayanan, and B. Dariush, “Spatio-temporal pyramid graph convolutions for human action recognition and postural assessment,” in IEEE/CVF Winter Conference on Applications of Computer Vision. IEEE, 2020, pp. 1069–1079. [Online]. Available: https://doi.org/10.1109/WACV45572.2020.9093368

    arXiv 2020

  10. [10]

    Openpose: Realtime multi-person 2d pose estimation using part affinity fields,

    Z. Cao, G. Hidalgo Martinez, T. Simon, S. Wei, and Y. A. Sheikh, “Openpose: Realtime multi-person 2d pose estimation using part affinity fields,” IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019

    2019

  11. [11]

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

    A. Shahroudy, J. Liu, T.-T. Ng, and G. Wang, “Ntu rgb+d: A large scale dataset for 3d human activity analysis,” in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 1010–1019

    2016

  12. [13]

    Skeleton based human action recognition for smart city application using deep learning,

    M. Rashmi and R. M. R. Guddeti, “Skeleton based human action recognition for smart city application using deep learning,” in 2020 international conference on communication systems & networks (COMSNETS). IEEE, 2020, pp. 756–761

    2020

  13. [14]

    Application of skeleton data and long short-term memory in action recognition of children with autism spectrum disorder,

    Y. Zhang, Y. Tian, P . Wu, and D. Chen, “Application of skeleton data and long short-term memory in action recognition of children with autism spectrum disorder,” Sensors, vol. 21, no. 2, p. 411,

  14. [15]

    A study on skeleton- based action recognition and its application to physical exercise recognition,

    Q. Pham, D. Nguyen, T. Nguyen, T. N. Nguyen, D. Nguyen, D. Pham, T. Tran, T. Le, and H. Vu, “A study on skeleton- based action recognition and its application to physical exercise recognition,” in The 11th International Symposium on Information and Communication Technology, SoICT 2022, Hanoi, Vietnam, December 1-3, 2022. ACM, 2022, pp. 239–246. [Online]. ...

    arXiv 2022

  15. [16]

    Skeleton-based action recognition with temporal action graph and temporal adaptive graph convolution structure,

    Y. Cao, C. Liu, Z. Huang, Y. Sheng, and Y. Ju, “Skeleton-based action recognition with temporal action graph and temporal adaptive graph convolution structure,” Multimedia Tools and Applications, vol. 80, no. 19, pp. 29 139–29 162, 2021. [Online]. Available: https://doi.org/10.1007/s11042-021-11136-z

    work page doi:10.1007/s11042-021-11136-z 2021

  16. [17]

    View transfer on human skeleton pose: Automatically disentangle the view-variant and view- invariant information for pose representation learning,

    Q. Nie and Y. Liu, “View transfer on human skeleton pose: Automatically disentangle the view-variant and view- invariant information for pose representation learning,” IJCV, vol. 129, no. 1, pp. 1–22, 2021. [Online]. Available: https: //doi.org/10.1007/s11263-020-01354-7

    work page doi:10.1007/s11263-020-01354-7 2021

  17. [18]

    Actional-structural graph convolutional networks for skeleton-based action recognition,

    M. Li, S. Chen, X. Chen, Y. Zhang, Y. Wang, and Q. Tian, “Actional-structural graph convolutional networks for skeleton-based action recognition,” in CVPR. Computer Vision Foundation / IEEE, 2019, pp. 3595–3603. [Online]. Available: http://openaccess.thecvf.com/content_CVPR_2019/ html/Li_Actional-Structural_Graph_Convolutional_Networks_ for_Skeleton-Based...

    2019

  18. [19]

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

    L. Shi, Y. Zhang, J. Cheng, and H. Lu, “Two-stream adaptive graph convolutional networks for skeleton-based action recognition,” in CVPR. Computer Vision Foundation / IEEE, 2019, pp. 12 026–12 035. [Online]. Available: http://openaccess.thecvf.com/content_CVPR_2019/html/Shi_ Two-Stream_Adaptive_Graph_Convolutional_Networks_for_ Skeleton-Based_Action_Recog...

    2019

  19. [20]

    Skeleton-based action recognition with directed graph neural networks,

    ——, “Skeleton-based action recognition with directed graph neural networks,” in CVPR. Computer Vision Foundation / IEEE, 2019, pp. 7912–7921. [Online]. Available: http://openaccess.thecvf.com/content_CVPR_2019/html/Shi_ Skeleton-Based_Action_Recognition_With_Directed_Graph_ Neural_Networks_CVPR_2019_paper.html

    2019

  20. [21]

    Skeleton- based action recognition with shift graph convolutional network,

    K. Cheng, Y. Zhang, X. He, W. Chen, J. Cheng, and H. Lu, “Skeleton- based action recognition with shift graph convolutional network,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 180–189

    2020

  21. [22]

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

    J. Liu, A. Shahroudy, M. Perez, G. Wang, L.-Y. Duan, and A. C. Kot, “Ntu rgb+ d 120: A large-scale benchmark for 3d human activity understanding,” IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 10, pp. 2684–2701, 2019

    2019

  22. [23]

    Fine-grained action retrieval through multiple parts-of-speech embeddings,

    M. Wray, G. Csurka, D. Larlus, and D. Damen, “Fine-grained action retrieval through multiple parts-of-speech embeddings,” in 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 450–459

    2019

  23. [24]

    Syntactically guided generative embeddings for zero-shot skeleton action recognition,

    P . Gupta, D. Sharma, and R. K. Sarvadevabhatla, “Syntactically guided generative embeddings for zero-shot skeleton action recognition,” in 2021 IEEE International Conference on Image Processing (ICIP), 2021, pp. 439–443

    2021

  24. [26]

    Sa-dvae: Improving zero-shot skeleton-based action recognition by disentangled variational autoencoders,

    S.-W. Li, Z.-X. Wei, W.-J. Chen, Y.-H. Yu, C.-Y. Yang, and J. Y. jen Hsu, “Sa-dvae: Improving zero-shot skeleton-based action recognition by disentangled variational autoencoders,” 2024. [Online]. Available: https://arxiv.org/abs/2407.13460

    Pith/arXiv arXiv 2024

  25. [27]

    Fine-grained side information guided dual-prompts for zero-shot skeleton action recognition,

    Y. Chen, J. Guo, T. He, and L. Wang, “Fine-grained side information guided dual-prompts for zero-shot skeleton action recognition,”

  26. [28]

    Language models are few-shot learners,

    T. B. Brown, B. Mann, N. Ryder, M. Subbiah, J. Kaplan, P . Dhariwal, A. Neelakantan, P . Shyam, G. Sastry, A. Askell, S. Agarwal, A. Herbert-Voss, G. Krueger, T. Henighan, R. Child, A. Ramesh, D. M. Ziegler, J. Wu, C. Winter, C. Hesse, M. Chen, E. Sigler, M. Litwin, S. Gray, B. Chess, J. Clark, C. Berner, S. McCandlish, A. Radford, I. Sutskever, and D. Am...

    2020

  27. [29]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agar- wal, G. Sastry, A. Askell, P . Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” 2021

    2021

  28. [30]

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

    L. Chunhui, H. Yueyu, L. Yanghao, S. Sijie, and L. Jiaying, “Pku- mmd: A large scale benchmark for continuous multi-modal human action understanding,” arXiv preprint arXiv:1703.07475, 2017

    Pith/arXiv arXiv 2017

  29. [31]

    Part-aware unified representation of language and skeleton for zero- shot action recognition,

    A. Zhu, Q. Ke, M. Gong, and J. Bailey, “Part-aware unified representation of language and skeleton for zero- shot action recognition,” 2024. [Online]. Available: https: //arxiv.org/abs/2406.13327

    Pith/arXiv arXiv 2024

  30. [32]

    Learning a deep embedding model for zero-shot learning,

    L. Zhang, T. Xiang, and S. Gong, “Learning a deep embedding model for zero-shot learning,” 2019

    2019

  31. [33]

    Hubness and pollu- tion: Delving into cross-space mapping for zero-shot learning,

    A. Lazaridou, G. Dinu, and M. Baroni, “Hubness and pollu- tion: Delving into cross-space mapping for zero-shot learning,” in Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). Beijing, China: Association for Computational...

    2015

  32. [34]

    Improving zero-shot learning by mitigating the hubness problem,

    G. Dinu, A. Lazaridou, and M. Baroni, “Improving zero-shot learning by mitigating the hubness problem,” 2015

    2015

  33. [35]

    Semantic autoencoder for zero-shot learning,

    E. Kodirov, T. Xiang, and S. Gong, “Semantic autoencoder for zero-shot learning,” 2017

    2017

  34. [36]

    Semantics disentangling for generalized zero-shot learning,

    Z. Chen, Y. Luo, R. Qiu, S. Wang, Z. Huang, J. Li, and Z. Zhang, “Semantics disentangling for generalized zero-shot learning,” 2021

    2021

  35. [37]

    Msdn: Mutually semantic distillation network for zero-shot learning,

    S. Chen, Z. Hong, G.-S. Xie, W. Yang, Q. Peng, K. Wang, J. Zhao, and X. You, “Msdn: Mutually semantic distillation network for zero-shot learning,” 2022

    2022

  36. [38]

    Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding,

    L. Xue, M. Gao, C. Xing, R. Martín-Martín, J. Wu, C. Xiong, R. Xu, J. C. Niebles, and S. Savarese, “Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding,” 2023

    2023

  37. [39]

    Pointclip: Point cloud understanding by clip,

    R. Zhang, Z. Guo, W. Zhang, K. Li, X. Miao, B. Cui, Y. Qiao, P . Gao, and H. Li, “Pointclip: Point cloud understanding by clip,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 8552–8562

    2022

  38. [40]

    Fine-grained action retrieval through multiple parts-of-speech embeddings,

    M. Wray, D. Larlus, G. Csurka, and D. Damen, “Fine-grained action retrieval through multiple parts-of-speech embeddings,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 450–459

    2019

  39. [41]

    Semantic-guided cross-modal prompt learning for skeleton-based zero-shot action recognition,

    A. Zhu, J. Zhu, J. Bailey, M. Gong, and Q. Ke, “Semantic-guided cross-modal prompt learning for skeleton-based zero-shot action recognition,” in Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), June 2025, pp. 13 876–13 885

    2025

  40. [42]

    Neuron: Learning context-aware evolving representations for zero-shot skeleton action recognition,

    Y. Chen, J. Guo, S. Guo, and D. Tao, “Neuron: Learning context-aware evolving representations for zero-shot skeleton action recognition,” 2024. [Online]. Available: https://arxiv.org/ abs/2411.11288

    Pith/arXiv arXiv 2024

  41. [43]

    Efficient test-time model adaptation without forgetting,

    S. Niu, J. Wu, Y. Zhang, Y. Chen, S. Zheng, P . Zhao, and M. Tan, “Efficient test-time model adaptation without forgetting,” in International conference on machine learning. PMLR, 2022, pp. 16 888–16 905

    2022

  42. [44]

    Tent: Fully test-time adaptation by entropy minimization,

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell, “Tent: Fully test-time adaptation by entropy minimization,” arXiv preprint arXiv:2006.10726, 2020

    Pith/arXiv arXiv 2006

  43. [45]

    Note: Robust continual test-time adaptation against temporal correlation,

    T. Gong, J. Jeong, T. Kim, Y. Kim, J. Shin, and S.-J. Lee, “Note: Robust continual test-time adaptation against temporal correlation,” Advances in Neural Information Processing Systems, vol. 35, pp. 27 253–27 266, 2022

    2022

  44. [46]

    Continual test-time domain adaptation,

    Q. Wang, O. Fink, L. Van Gool, and D. Dai, “Continual test-time domain adaptation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 7201–7211

    2022

  45. [47]

    Adanpc: Exploring non-parametric classifier for test- time adaptation,

    Y. Zhang, X. Wang, K. Jin, K. Yuan, Z. Zhang, L. Wang, R. Jin, and T. Tan, “Adanpc: Exploring non-parametric classifier for test- time adaptation,” in International conference on machine learning. PMLR, 2023, pp. 41 647–41 676

    2023

  46. [48]

    Test-time prompt tuning for zero-shot generalization in vision-language models,

    M. Shu, W. Nie, D.-A. Huang, Z. Yu, T. Goldstein, A. Anandkumar, and C. Xiao, “Test-time prompt tuning for zero-shot generalization in vision-language models,” Advances in Neural Information Processing Systems, vol. 35, pp. 14 274–14 289, 2022

    2022

  47. [49]

    Diverse data augmentation with diffusions for effective test-time prompt tuning,

    C.-M. Feng, K. Yu, Y. Liu, S. Khan, and W. Zuo, “Diverse data augmentation with diffusions for effective test-time prompt tuning,”

  48. [50]

    Calip: Zero-shot enhancement of clip with parameter-free attention,

    Z. Guo, R. Zhang, L. Qiu, X. Ma, X. Miao, X. He, and B. Cui, “Calip: Zero-shot enhancement of clip with parameter-free attention,”

  49. [51]

    Efficient test-time adaptation of vision-language models,

    A. Karmanov, D. Guan, S. Lu, A. El Saddik, and E. Xing, “Efficient test-time adaptation of vision-language models,” in Proceedings 16 of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 14 162–14 171

    2024

  50. [52]

    A review of generalized zero-shot learning methods,

    F. Pourpanah, M. Abdar, Y. Luo, X. Zhou, R. Wang, C. P . Lim, X.-Z. Wang, and Q. J. Wu, “A review of generalized zero-shot learning methods,” IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 4, pp. 4051–4070, 2022

    2022

  51. [53]

    Memo: Test time robustness via adaptation and augmentation,

    M. Zhang, S. Levine, and C. Finn, “Memo: Test time robustness via adaptation and augmentation,” 2022. [Online]. Available: https://arxiv.org/abs/2110.09506

    Pith/arXiv arXiv 2022

  52. [55]

    Discriminability- driven channel selection for out-of-distribution detection,

    Y. Yuan, R. He, Y. Dong, Z. Han, and Y. Yin, “Discriminability- driven channel selection for out-of-distribution detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26 171–26 180

    2024

  53. [56]

    Unified entropy optimization for open-set test-time adaptation,

    Z. Gao, X.-Y. Zhang, and C.-L. Liu, “Unified entropy optimization for open-set test-time adaptation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2024, pp. 23 975–23 984

    2024

  54. [57]

    Entropy maximization and meta classification for out-of-distribution detection in seman- tic segmentation,

    R. Chan, M. Rottmann, and H. Gottschalk, “Entropy maximization and meta classification for out-of-distribution detection in seman- tic segmentation,” in Proceedings of the ieee/cvf international conference on computer vision, 2021, pp. 5128–5137

    2021

  55. [58]

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

    A. Shahroudy, J. Liu, T.-T. Ng, and G. Wang, “Ntu rgb+ d: A large scale dataset for 3d human activity analysis,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 1010–1019

    2016

  56. [59]

    Syntactically guided generative embeddings for zero-shot skeleton action recognition,

    P . Gupta, D. Sharma, and R. K. Sarvadevabhatla, “Syntactically guided generative embeddings for zero-shot skeleton action recognition,” 2021. [Online]. Available: https://arxiv.org/abs/2101. 11530

    2021

  57. [60]

    Zero- shot skeleton-based action recognition via mutual information estimation and maximization,

    Y. Zhou, W. Qiang, A. Rao, N. Lin, B. Su, and J. Wang, “Zero- shot skeleton-based action recognition via mutual information estimation and maximization,” in Proceedings of the 31st ACM International Conference on Multimedia, ser. MM ’23. ACM, Oct. 2023, p. 5302–5310. [Online]. Available: http://dx.doi.org/10.1145/3581783.3611888

    arXiv 2023

  58. [61]

    Devise: A deep visual-semantic embedding model,

    A. Frome, G. S. Corrado, J. Shlens, S. Bengio, J. Dean, M. A. Ranzato, and T. Mikolov, “Devise: A deep visual-semantic embedding model,” in Advances in Neural Information Processing Systems, C. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K. Weinberger, Eds., vol. 26. Curran Associates, Inc., 2013. [Online]. Available: https://proceedings.neurips.cc/...

    2013

  59. [62]

    Learning robust visual-semantic embeddings,

    Y.-H. Hubert Tsai, L.-K. Huang, and R. Salakhutdinov, “Learning robust visual-semantic embeddings,” in Proceedings of the IEEE International conference on Computer Vision, 2017, pp. 3571–3580

    2017

  60. [63]

    Gen- eralized zero-shot learning via aligned variational autoencoders,

    E. Schonfeld, S. Ebrahimi, S. Sinha, T. Darrell, and Z. Akata, “Gen- eralized zero-shot learning via aligned variational autoencoders,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2019, pp. 54–57

    2019

  61. [64]

    Skeleton- based action recognition with shift graph convolutional network,

    K. Cheng, Y. Zhang, X. He, W. Chen, J. Cheng, and H. Lu, “Skeleton- based action recognition with shift graph convolutional network,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 183–192

    2020

  62. [65]

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

    S. Yan, Y. Xiong, and D. Lin, “Spatial temporal graph convolutional networks for skeleton-based action recognition,” 2018

    2018

  63. [66]

    Skeleton-based action recog- nition with multi-stream adaptive graph convolutional networks,

    L. Shi, Y. Zhang, J. Cheng, and H. Lu, “Skeleton-based action recog- nition with multi-stream adaptive graph convolutional networks,” IEEE Transactions on Image Processing, vol. 29, pp. 9532–9545, 2020

    2020

  64. [67]

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

    Y. Chen, Z. Zhang, C. Yuan, B. Li, Y. Deng, and W. Hu, “Channel- wise topology refinement graph convolution for skeleton-based action recognition,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 13 359–13 368

    2021

  65. [68]

    Revisiting skeleton- based action recognition,

    H. Duan, Y. Zhao, K. Chen, D. Lin, and B. Dai, “Revisiting skeleton- based action recognition,” arXiv preprint arXiv:2104.13586, 2021. Jingmin Zhureceived his Bachelor’s degree in Physics from Southern University of Science and Technology, Shenzhen, China, in July 2023, and his Master’s degree in Artificial Intelligence from Monash University, Melbourne, ...

    Pith/arXiv arXiv 2021

  66. [2018]

    Available: http://arxiv.org/abs/1807.03748

    [Online]. Available: http://arxiv.org/abs/1807.03748

    Pith/arXiv arXiv

  67. [2021]

    Available: https://doi.org/10.3390/s21020411 15

    [Online]. Available: https://doi.org/10.3390/s21020411 15

    work page doi:10.3390/s21020411

  68. [2022]

    Available: https://arxiv.org/abs/2209.14169

    [Online]. Available: https://arxiv.org/abs/2209.14169

    Pith/arXiv arXiv

  69. [2023]

    Available: https://arxiv.org/abs/2308.06038

    [Online]. Available: https://arxiv.org/abs/2308.06038

    Pith/arXiv arXiv

  70. [2024]

    Available: https://arxiv.org/abs/2404.07487

    [Online]. Available: https://arxiv.org/abs/2404.07487

    Pith/arXiv arXiv