Pith. sign in

REVIEW 3 major objections 6 minor 37 references

Skeleton-Cache claims that zero-shot skeleton action recognition can be improved by 4-8 percentage points at test time with no training, by retrieving and LLM-weighting structured skeleton descriptors from a growing cache.

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 16:52 UTC pith:5LCV7GLW

load-bearing objection Useful training-free TTA for skeleton ZSL with consistent gains, but the self-retrieval confound and test-set hyperparameter selection need attention before the exact numbers are trusted. the 3 major comments →

arxiv 2512.11458 v1 pith:5LCV7GLW submitted 2025-12-12 cs.CV cs.AI

Boosting Skeleton-based Zero-Shot Action Recognition with Training-Free Test-Time Adaptation

Jingmin Zhu , Anqi Zhu , Hossein Rahmani , Jun Liu , Mohammed Bennamoun , Qiuhong Ke This is my paper
classification cs.CV cs.AI
keywords skeleton-based zero-shot action recognitiontest-time adaptationtraining-freenon-parametric cacheretrievalLLM-guided fusiongeneralized zero-shot learninghuman action recognition
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.

Skeleton-Cache claims that a frozen skeleton-based zero-shot action-recognition model can recognize unseen actions substantially better during inference without any gradient updates, retraining, or access to training data. The method splits each test sequence into eight descriptors — one global, four body-part, three temporal-phase — and stores confident samples in a non-parametric cache organized by predicted class. At query time, it retrieves descriptor-wise class logits from the cache and fuses them with class-specific weights that a large language model generates from the action name alone. On NTU RGB+D 60/120 and PKU-MMD II the module adds 4-8 percentage points of accuracy across several existing backbones, in both zero-shot and generalized zero-shot protocols. If the claim holds, it gives practitioners a cheap plug-in way to adapt skeleton recognizers to novel actions at deployment.

Core claim

The paper's central claim is that test-time adaptation for skeleton-based zero-shot action recognition can be done without training by turning inference into retrieval over a growing non-parametric cache. Each entry stores a structured key — one global, four body-part, three temporal-phase descriptors from the frozen encoder — plus the model's predicted label and an entropy confidence. High-confidence samples enter their class block; low-confidence entries are replaced. A query is matched descriptor-by-descriptor via temperature-scaled cosine affinity, producing per-descriptor class logits, which an LLM-generated per-class weight vector fuses. The adapted logits are the frozen model's logits

What carries the argument

The load-bearing mechanism is a structured non-parametric cache with descriptor-wise retrieval and LLM-guided fusion. The cache key for a skeleton sequence is a matrix formed by concatenating one globally pooled feature, four body-part averaged features (head, torso, arms, feet), and three temporal-segment features (begin, middle, end). Affinity between a query descriptor and cached keys is the exponential of a negative cosine distance scaled by a temperature; these affinities are projected through a one-hot label matrix into per-descriptor class logits. A single prompt to a large language model per action class returns normalized spatial, temporal, and global-vs-local weights, which decide

Load-bearing premise

The load-bearing premise is that the frozen model's high-confidence pseudo-labels are mostly correct for each class, so cache entries store representative patterns rather than systematic errors; if that fails, retrieval reinforces the model's mistakes instead of correcting them.

What would settle it

On a fixed split such as NTU60 48/12, compute the precision of cached entries for the weakest backbone (what fraction of entries in each class block actually carry the correct label). Then replace the pseudo-label filter with an oracle labeler on the same cache update rule. If oracle labels do not raise accuracy above the pseudo-label version, or if a cache filled with random high-confidence entries matches the method's 4-8 point gain, the claimed mechanism — selective caching of confident, accurate exemplars — would be falsified.

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

If this is right

  • Any frozen SZAR backbone can be boosted by attaching the plug-in cache; reported gains are 4-8 percentage points across models, datasets, and splits.
  • Because no gradients are used, adaptation runs at streaming rates with under 3 MB extra memory and modest added latency, supporting real-time deployment.
  • The method improves generalized zero-shot performance (seen/unseen harmonic mean by 4-7 points), so it does not sacrifice already-known classes while adapting to unseen ones.
  • LLM-derived descriptor weights let the system exploit commonsense action semantics — which body parts and motion phases matter — without any labels for unseen classes.
  • The cache's memory and latency scale with the number of unseen classes and cache size, not with sequence length, keeping the overhead predictable.

Where Pith is reading between the lines

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

  • A stress test follows from the paper's own numbers: on NTU60 48/12 the weakest backbone starts below 41% base accuracy, so a large fraction of cache entries are likely mislabeled; the reported gain then tests whether even an imperfect cache can help retrieval. Measuring cached-entry label precision directly would show whether the mechanism is pseudo-label accuracy or structural descriptor consiste
  • If the benefit comes mainly from consistent local descriptors rather than correct labels, a natural extension is to replace hard one-hot cache values with soft or entropy-weighted labels, which could further stabilize weak backbones.
  • Because the LLM weights are generated once per dataset and frozen, the method inherits any LLM bias about which body parts matter; per-sample or per-domain adaptive weights are a testable upgrade not explored in the paper.
  • The global/local-spatial/temporal decomposition is not skeleton-specific in principle; applying the same cache-plus-LLM-weight recipe to RGB video features or other structured modalities is a plausible extension the paper does not make.

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 / 6 minor

Summary. The paper proposes Skeleton-Cache, a training-free test-time adaptation (TTA) module for skeleton-based zero-shot action recognition (SZAR). The method builds a non-parametric cache of structured skeleton descriptors (one global, four body-part, and three temporal-phase descriptors) that is updated online with pseudo-labels and confidence scores from a frozen SZAR backbone. At test time, descriptor-wise retrieval logits are computed by cosine affinity against the cache, fused with class-specific weights obtained from an LLM (GPT-4o), and combined with the frozen model's original logits via a balancing coefficient. Experiments on NTU RGB+D 60/120 and PKU-MMD II report consistent 4–8 percentage-point improvements across multiple SZAR backbones under both ZSL and GZSL settings, together with ablations on cache size, fusion weights, descriptor granularity, and comparisons with other TTA methods.

Significance. If the reported gains are genuine, Skeleton-Cache would be a useful practical contribution: it is plug-and-play, training-free, and requires no source data, while leveraging structured skeletal cues and LLM priors in a way that is new to SZAR. The paper is clearly written, the algorithm is specified in detail, and the public code will aid reproducibility. The ablation of LLM-guided weights against uniform and random weights is a strength, as is the inclusion of efficiency measurements. However, the central claim that the gains come from correct pseudo-label propagation through the cache is not yet established, because the current protocol has a self-retrieval confound and the cache-insertion mechanism is not ablated.

major comments (3)
  1. [Algorithm 1 and Section 3.2 (Eqs. 6, 8, 11, 12)] The cache is updated with the current test sample before retrieval, so every query retrieves against a cache that contains itself. In Eq. 6, the query's own cached key has cosine similarity 1, giving affinity 1 for the predicted class in every descriptor. After the LLM-weighted fusion (Eq. 11) and scaling by α_s=5.0 (Eq. 12), the predicted class receives an added logit of α_s·1 = 5.0, independent of whether the pseudo-label is correct. This is a large, unablated inductive bias toward the frozen model's prediction, and it is absent from the inductive and TTA baselines. The paper does not report results with the cache updated after retrieval or with the self-entry excluded, so the causal claim that the 4–8 pp improvements come from correct pseudo-label propagation is not established. Please add this ablation and quantify the contribution of self-retrieval.
  2. [Section 4.5, Fig. 2, and Tables 1–4] The hyperparameters K, α_s, and β are selected by ablations on the same test splits used for the final reported numbers, with no held-out validation. The paper states performance is 'robust' to hyperparameter choices, but all main results use the peak of the test-split curves. Moreover, all accuracy numbers appear to come from single runs with no error bars or significance tests. Given the reported gains sometimes range from 3.5 to 7.2 pp, the exact magnitude of the improvement is uncertain. Please provide nested validation (or a clear statement of protocol), report variance across multiple runs, or at least show that the conclusions are stable across a range of hyperparameters.
  3. [Section 3.2 (Cache Update) and Algorithm 1] The cache stores pseudo-labels from a frozen model, and for weak backbones (e.g., PURLS at 40.99% on NTU60 48/12) a majority of cached entries will be mislabeled. The 'low-confidence replacement' rule is not a threshold: when a class block is not full, samples are inserted regardless of confidence; only when the block is full is the highest-entropy entry replaced. Thus the cache can contain many arbitrary low-confidence, incorrectly labeled entries. Appendix C.1 shows that stronger backbones gain more, which is consistent with the pseudo-label-quality hypothesis, but the paper never measures the accuracy of cached entries or of the low-entropy subset. To support the mechanism, please report oracle-label cache accuracy and/or the precision of cached entries, and ablate the entropy-based replacement against random replacement.
minor comments (6)
  1. [Eq. (3)] The three descriptor formulas are typeset in a garbled way: 'sp = ... t z = ... g=...' is visually confusing and appears to combine three equations into one line. Please separate the definitions of s_p, t_z, and g.
  2. [Section 4.3] The GZSL metric description says 'accuracy on both seen test samples D^u_te and unseen test samples D^u_te'; the seen and unseen sets are given the same symbol. Use D^s_te and D^u_te.
  3. [Algorithm 1 and Section 3.2] The cache entry is described as a tuple (e_k, e_y, e_h), but Algorithm 1 stores only (e_k, h) in the class block, with the label determined by the block index. To match the text, either store the label explicitly or clarify that e_y is implicit in the block assignment.
  4. [Table 3 (Weighting strategy)] The 'Random weights' row appears to be a single random draw. Since random weights have high variance, please report the mean and standard deviation over multiple random seeds.
  5. [Appendix B.2] The main text says GPT-4o is used for weight generation, while Appendix B.2 says 'gpt-4-turbo with temperature τ=0'. Please reconcile the exact model and version, since the LLM output is part of the method.
  6. [Appendix C.3, Table 8] The text reports '44.96%' for SC (Adapted Logits) on NTU60 48/12, but Table 8 shows '44.66'. Please correct the inconsistency.

Circularity Check

0 steps flagged

No significant circularity: the method is a transductive, training-free cache that blends base-model logits with descriptor similarities and external LLM priors; reported gains are benchmarked externally and no prediction reduces by construction to a fitted input.

full rationale

The paper's derivation chain is self-contained and empirical rather than a claimed first-principles result. Skeleton-Cache extracts global/spatial/temporal descriptors from a frozen SZAR encoder (Eq. 3), updates a cache with the frozen model's own pseudo-labels filtered by entropy (Section 3.2, Algorithm 1), computes affinity-weighted descriptor-wise logits (Eqs. 6-8), fuses them using LLM-derived class weights (Eqs. 9-11), and adds the result to the original zero-shot logits (Eq. 12). The cache pseudo-labels are inputs to a feedback loop common to test-time adaptation, but the final prediction is not equal to the pseudo-label by construction: it remains a weighted combination of base logits, many cached entries, and descriptor similarities, and the LLM weights are generated independently from action names. The self-referential update (a query is inserted before retrieval, giving a self-similarity contribution to its own predicted class) is a design detail that could be ablated, but it is not an equation-level reduction of the output to a fitted quantity, and it cannot by itself explain the reported 4-8 pp gains across heterogeneous backbones. The only self-citation is PURLS, used as one of several backbones rather than as a load-bearing theoretical premise; no uniqueness theorem or ansatz is imported from the authors' prior work. Appendix C.1 explicitly acknowledges that weaker backbones benefit less because their predictions are less confident, which is a limitation of the pseudo-label premise, not a circular justification. The evaluations use external benchmarks (NTU RGB+D, PKU-MMD), and no step in the paper equates a prediction with a fitted parameter or a self-cited result. Therefore no significant circularity is present.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central model is a retrieval system; it uses no new physical entities. It does introduce several hand-chosen hyperparameters tuned on the same test splits, and it relies on three domain assumptions about skeleton features, pseudo-label quality, and LLM priors.

free parameters (3)
  • cache size K = 8
    Ablated on the same four test splits in Fig. 2a; no held-out validation reported.
  • fusion coefficient alpha_s = 5.0
    Ablated on the same four test splits in Fig. 2b; optimal value chosen from test accuracy.
  • similarity temperature beta = 3.0
    Ablated on the same four test splits in Fig. 2c; optimal value chosen from test accuracy.
axioms (4)
  • domain assumption Averaging the frozen ST-GCN latent tensor over joint groups and time segments retains enough discriminative signal for retrieval (Eq. 3-4).
    The whole method assumes global, spatial, and temporal averages are informative descriptors for distinguishing unseen actions; this is not proven, only empirically supported.
  • domain assumption The frozen model's pseudo-labels, filtered by low entropy, are reliable enough to populate the cache.
    Algorithm 1 stores entries under the predicted class y-hat; if these pseudo-labels are mostly wrong, the retrieval logits may reinforce errors.
  • domain assumption LLM-generated class-specific weights reflect which body parts and temporal phases matter for recognizing the action from skeletons.
    The LLM is queried with action names only; the paper assumes its commonsense priors transfer to skeleton motion features.
  • standard math Cosine similarity with an exponential kernel (Eq. 6) is a suitable affinity measure for skeleton descriptors.
    This is a standard retrieval kernel, not an ad hoc invention, but its adequacy is an assumption.

pith-pipeline@v1.3.0-alltime-deepseek · 19068 in / 13803 out tokens · 126818 ms · 2026-08-03T16:52:17.429235+00:00 · methodology

0 comments
read the original abstract

We introduce Skeleton-Cache, the first training-free test-time adaptation framework for skeleton-based zero-shot action recognition (SZAR), aimed at improving model generalization to unseen actions during inference. Skeleton-Cache reformulates inference as a lightweight retrieval process over a non-parametric cache that stores structured skeleton representations, combining both global and fine-grained local descriptors. To guide the fusion of descriptor-wise predictions, we leverage the semantic reasoning capabilities of large language models (LLMs) to assign class-specific importance weights. By integrating these structured descriptors with LLM-guided semantic priors, Skeleton-Cache dynamically adapts to unseen actions without any additional training or access to training data. Extensive experiments on NTU RGB+D 60/120 and PKU-MMD II demonstrate that Skeleton-Cache consistently boosts the performance of various SZAR backbones under both zero-shot and generalized zero-shot settings. The code is publicly available at https://github.com/Alchemist0754/Skeleton-Cache.

Figures

Figures reproduced from arXiv: 2512.11458 by Anqi Zhu, Hossein Rahmani, Jingmin Zhu, Jun Liu, Mohammed Bennamoun, Qiuhong Ke.

Figure 1
Figure 1. Figure 1: (a) Overview of the proposed pipeline that integrates our Skeleton [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Ablation studies on key hyperparameters of the Skeleton-Cache mechanism. (a) Cache size K. (b) Balancing coefficient αs. (c) Similarity temperature parameter β [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Heat-map visualisation of GPT-4–derived weights w(c) . Columns correspond to the eight descriptors; rows correspond to unseen classes of NTU 55/5 split. G H T A L B M E Descriptors (G=Global, H=Head, T=Torso, A=Arms, L=Legs, B=Begin, M=Middle, E=End) drop put on jacket salute headache punch or slap juggle table tennis table put object into bag cross arms butt kicks wield knife Action Classes LLM-Prior Weig… view at source ↗
Figure 4
Figure 4. Figure 4: Heat-map visualisation of GPT-4–derived weights w(c) . Columns correspond to the eight descriptors; rows correspond to unseen classes of NTU 110/10 split. Figures 3 and 4 display heatmaps of the GPT-4-derived weights w(c) for the NTU 55/5 and 110/10 splits, respectively. These weights align closely with human intuition about actions. For example, in the NTU 55/5 split, reading and writing emphasize the arm… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of confusion matrices on NTU 55/5 split. Adaptation with Skeleton [PITH_FULL_IMAGE:figures/full_fig_p020_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of confusion matrices on NTU 110/10 split. Adaptation with Skeleton [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of Top-5 prediction changes before and after applying the Skeleton-Cache for [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-class accuracy comparison between base PURLS and adapted PURLS+SC for NTU [PITH_FULL_IMAGE:figures/full_fig_p023_8.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

37 extracted references · 1 canonical work pages

  1. [1]

    W. Cao, Y . Wu, Y . Sun, H. Zhang, J. Ren, D. Gu, and X. Wang. A review on multimodal zero-shot learning.Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery, 13(2):e1488, 2023

    2023

  2. [2]

    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, 80(19):29139–29162, 2021

    2021

  3. [3]

    Y . Chen, J. Guo, T. He, and L. Wang. Fine-grained side information guided dual-prompts for zero-shot skeleton action recognition, 2024. URLhttps://arxiv.org/abs/2404.07487

    Pith/arXiv arXiv 2024

  4. [4]

    Chunhui, H

    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

  5. [5]

    H. Duan, Y . Zhao, K. Chen, D. Lin, and B. Dai. Revisiting skeleton-based action recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2969–2978, 2022

    2022

  6. [6]

    C.-M. Feng, K. Yu, Y . Liu, S. Khan, and W. Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning, 2023. URLhttps://arxiv.org/abs/2308.06038

    Pith/arXiv arXiv 2023

  7. [7]

    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, 2022. URLhttps://arxiv.org/abs/2209.14169

    Pith/arXiv arXiv 2022

  8. [8]

    Gupta, D

    P. Gupta, D. Sharma, and R. K. Sarvadevabhatla. Syntactically guided generative embeddings for zero-shot skeleton action recognition, 2021. URLhttps://arxiv.org/abs/2101.11530

    Pith/arXiv arXiv 2021

  9. [9]

    Hubert Tsai, L.-K

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

    2017

  10. [10]

    Iwasawa and Y

    Y . Iwasawa and Y . Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization.Advances in Neural Information Processing Systems, 34:2427–2440, 2021

    2021

  11. [11]

    Karmanov, D

    A. Karmanov, D. Guan, S. Lu, A. El Saddik, and E. Xing. Efficient test-time adaptation of vision-language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14162–14171, 2024

    2024

  12. [12]

    Karunaratne, A

    I. Karunaratne, A. S. Atukorale, and H. Perera. Surveillance of human-computer interactions: A way forward to detection of users’ psychological distress. In2011 IEEE Colloquium on Humanities, Science and Engineering, pages 491–496. IEEE, 2011

    2011

  13. [13]

    A. L. Kotian, R. Nandipi, U. M, U. R. S, V ARSHAUK, and V . G. T. A systematic review on human and computer interaction. In2024 2nd International Conference on Intelligent Data Communication Technologies and Internet of Things (IDCIoT), pages 1214–1218, 2024. doi: 10.1109/IDCIoT59759.2024.10467622

    arXiv 2024

  14. [14]

    M. Li, S. Chen, X. Chen, Y . Zhang, Y . Wang, and Q. Tian. Actional-structural graph con- volutional networks for skeleton-based action recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3595–3603, 2019

    2019

  15. [15]

    Li, Z.-X

    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. URLhttps://arxiv.org/abs/2407.13460. 11

    Pith/arXiv arXiv 2024

  16. [16]

    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, 42(10):2684–2701, 2019

    2019

  17. [17]

    Pourpanah, M

    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, 45(4):4051–4070, 2022

    2022

  18. [18]

    J. H. A. Samadh, H. Gani, N. H. Hussein, M. U. Khattak, M. Naseer, F. Khan, and S. Khan. Align your prompts: Test-time prompting with distribution alignment for zero-shot generalization. In Thirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  19. [19]

    Schonfeld, S

    E. Schonfeld, S. Ebrahimi, S. Sinha, T. Darrell, and Z. Akata. Generalized zero-shot learning via aligned variational autoencoders. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 54–57, 2019

    2019

  20. [20]

    Shahroudy, J

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

    2016

  21. [21]

    M. B. Shaikh, D. Chai, S. M. S. Islam, and N. Akhtar. From cnns to transformers in multimodal human action recognition: A survey.ACM Transactions on Multimedia Computing, Commu- nications, and Applications, 20(8):1–24, July 2024. ISSN 1551-6865. doi: 10.1145/3664815. URLhttp://dx.doi.org/10.1145/3664815

    work page doi:10.1145/3664815 2024

  22. [22]

    L. Shi, Y . Zhang, J. Cheng, and H. Lu. Skeleton-based action recognition with directed graph neural networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7912–7921, 2019

    2019

  23. [23]

    L. Shi, Y . Zhang, J. Cheng, and H. Lu. Two-stream adaptive graph convolutional networks for skeleton-based action recognition. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12026–12035, 2019

    2019

  24. [24]

    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, 35:14274–14289, 2022

    2022

  25. [25]

    Z. Tang, Y . Zhao, Y . Wen, and M. Liu. A survey on backbones for deep video action recognition,

  26. [26]

    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

  27. [27]

    Z. Wang, Y . Yang, Z. Liu, and Y . Zheng. Deep neural networks in video human action recognition: A review, 2023. URLhttps://arxiv.org/abs/2305.15692

    Pith/arXiv arXiv 2023

  28. [28]

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

    2019

  29. [29]

    D. Yang, Y . Wang, A. Dantcheva, L. Garattoni, G. Francesca, and F. Brémond. View-invariant skeleton action representation learning via motion retargeting.International Journal of Com- puter Vision, pages 1–16, 2024

    2024

  30. [30]

    B. Yu, H. Yin, and Z. Zhu. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. InProceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence, IJCAI-2018, page 3634–3640. International Joint Confer- ences on Artificial Intelligence Organization, July 2018. doi: 10.24963/ijcai.2018...

    doi:10.24963/ijcai.2018/505 2018

  31. [31]

    Zhang, X

    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. InInternational conference on machine learning, pages 41647–41676. PMLR, 2023. 12

    2023

  32. [32]

    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. InProceedings of the 31st ACM International Conference on Multimedia, MM ’23, page 5302–5310. ACM, Oct. 2023. doi: 10.1145/3581783.3611888. URLhttp://dx.doi.org/10.1145/3581783.3611888

    arXiv 2023

  33. [33]

    A. Zhu, Q. Ke, M. Gong, and J. Bailey. Part-aware unified representation of language and skele- ton for zero-shot action recognition, 2024. URLhttps://arxiv.org/abs/2406.13327. 13 Appendix This appendix provides detailed supplementary material to support the main paper,Boosting Skeleton- based Zero-Shot Action Recognition with Training-Free Test-Time Adap...

    Pith/arXiv arXiv 2024

  34. [35]

    Provide a list of four non-negative numbers that sum to 1, corresponding to the relative importance of each region for recognising<ACTION>

    Spatial importance.The human body is divided into four regions: [Head, Torso, Arms, Legs] . Provide a list of four non-negative numbers that sum to 1, corresponding to the relative importance of each region for recognising<ACTION>. Format:"spatial": [w_head, w_torso, w_arms, w_legs]

  35. [36]

    temporal

    Temporal importance.The action sequence is divided into three phases: [Beginning, Middle, End]. Provide a list of three non-negative numbers that sum to 1, indicating the relative importance of each phase. Format:"temporal": [w_begin, w_mid, w_end]

  36. [37]

    gamma":γ Returnone compact JSON objectwith keys

    Global vs local preference.Provide a single number γ∈[0,1] indicating how much the action should be recognised holistically (γ≈1) versus by local parts/phases (γ≈0). Format:"gamma":γ Returnone compact JSON objectwith keys "spatial", "temporal", and "gamma". Do not include any other keys, text, or explanations. 15 Example Response (action = "Waving") { "sp...

  37. [2024]

    URLhttps://arxiv.org/abs/2405.05584

    Pith/arXiv arXiv