Pith. sign in

REVIEW 3 major objections 5 minor 35 references

A pretrained temporal transformer rivals graph-based skeleton action models

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 ↗

A masked-pretrained skeleton transformer with a second fine-tuning transformer and cross-attention fusion reaches 94.66% on Penn Action, 91.16% on N-UCLA, and 81.01%/88.17% on NTU RGB+D 60 cross-subject/cross-view.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection CascadeFormer is a clean, honest empirical paper with released code, but its central claim—that the T2 cascade stage helps—is never actually tested. the 3 major comments →

arxiv 2509.00692 v1 pith:WFLSQ2VL submitted 2025-08-31 cs.CV

CascadeFormer: A Family of Two-stage Cascading Transformers for Skeleton-based Human Action Recognition

Yusen Peng , Alper Yilmaz This is my paper
classification cs.CV
keywords skeleton-based action recognitionmasked pretrainingcascading fine-tuningtemporal transformercross-attentionNTU RGB+D 60graph convolutional networksself-supervised learning
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.

The reading

CascadeFormer argues that a plain temporal transformer, trained first to reconstruct randomly masked joint coordinates and then refined by a second task-specific transformer, can match graph-convolutional models on skeleton-based action recognition without ever modeling the skeleton as a graph. On NTU RGB+D 60 cross-subject, the best variant reaches 81.01% accuracy against ST-GCN's 81.5%, and on cross-view 88.17% against 88.3%. The same two-stage pipeline reaches 94.66% on Penn Action and 91.16% on N-UCLA, transferring across 2D and 3D skeleton data. The claim matters because it suggests masked pretraining plus cascading fine-tuning can supply the spatial structure that graph methods build in by design, making a vanilla transformer a credible default architecture for skeleton action recognition.

Core claim

On its own terms, the paper's central claim is that joint-level masked pretraining—masking 30% of joints across all frames and reconstructing the masked coordinates with a lightweight linear decoder—gives a vanilla temporal transformer enough spatiotemporal structure that a subsequent cascading fine-tuning stage can compete with graph-convolutional baselines. The cascading stage passes the pretrained frame embeddings through a second task-specific transformer and fuses them with the original embeddings via cross-attention before frame-average pooling and classification. The decisive evidence is on NTU RGB+D 60: CascadeFormer 1.0 reaches 81.01% cross-subject and 88.17% cross-view, essentially

What carries the argument

The load-bearing mechanism is the two-stage cascade: a temporal transformer (T1) pretrained to reconstruct masked joint coordinates with a linear decoder, followed by a task-specific transformer (T2) that refines T1's frame embeddings; the refined and original embeddings are combined through cross-attention, average-pooled over frames, and classified. The three CascadeFormer variants share T1, T2, pretraining, and fine-tuning stages, differing only in how frame tokens are built—linear projection, a 1D convolution over joints, or a single spatial transformer over joint embeddings. What carries the argument is minimalism: no graph edges or skeleton topology appear anywhere, so competitive accu

Load-bearing premise

That the added second-stage transformer is what makes the fine-tuning stage work: the paper never compares against fine-tuning the same pretrained backbone with only a linear classifier head, so if that control matches the reported accuracy, the cascading novelty adds nothing.

What would settle it

Run the same 100-epoch masked pretraining on NTU RGB+D 60 cross-subject and fine-tune in two ways: with the cascading T2 transformer, and with only frame-average pooling plus a linear classifier. If the linear-head run reaches 81.01% or higher, the cascading stage is not the cause of the result. Separately, rerun Penn Action evaluation without removing occluded skeletons; if accuracy drops toward or below the 93.4% HDM-BG baseline, the comparison protocol changes the conclusion.

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

If this is right

  • On NTU RGB+D 60, CascadeFormer 1.0 reaches 81.01% cross-subject and 88.17% cross-view, showing that spatial graphs are not required to stay competitive with graph-convolutional baselines on large-scale skeleton data.
  • The same two-stage pipeline transfers across dataset formats and scales: 94.66% on Penn Action (2D, 13 joints), 91.16% on N-UCLA (3D, 20 joints), and strong results on NTU60 (3D, 25 joints).
  • The pretraining-duration ablation ties downstream accuracy directly to masked pretraining: 76.38% after 1 epoch versus 81.01% after 100 epochs on NTU60 cross-subject.
  • Joint-level random masking outperforms frame-level masking (94.66% vs 89.98%) and unmasked reconstruction (91.10%) on Penn Action, indicating that fine-grained spatial masking is the more effective pretraining objective.
  • Among the three feature extractors, the simplest linear frame projection is best or near-best on all datasets, so added spatial modeling inside the encoder is not what drives performance.

Where Pith is reading between the lines

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

  • If a future control run fine-tunes the same pretrained backbone with only a linear classifier head and matches the reported accuracy, the second transformer T2 would be shown to add little; the paper's reported gains do not settle that question directly.
  • The Penn Action evaluation removes occluded skeletons before training and testing, a protocol change the cited baselines are not described as using; rerunning on the original unmodified set would test whether the 94.66% ranking holds.
  • The cross-view gap on two-person actions (84.86% vs 88.92% for single-person actions) suggests that inter-person dynamics, not joint topology, is the next bottleneck; adding explicit relative-pose or interaction features between persons is a direct extension.
  • Because the simplest encoder wins on most splits, the results imply that frame-level tokens plus masked joint reconstruction can absorb most of the spatial information in skeleton data, so graph or part-based encoders may be replaceable in other skeleton pipelines as well.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper introduces CascadeFormer, a family of two-stage transformer models for skeleton-based action recognition. The first stage applies masked joint-level pretraining to a temporal transformer backbone (T1) with a linear reconstruction decoder. The second stage is a cascading fine-tuning component in which a task-specific transformer (T2) refines T1's frame embeddings, fuses them via cross-attention, and feeds the pooled result to a linear classifier. Three variants (1.0, 1.1, 1.2) differ only in the feature extraction module (linear projection, 1D convolution, and joint-level spatial transformer, respectively). The models are evaluated on Penn Action, N-UCLA, and NTU RGB+D 60 (cross-subject and cross-view), with reported accuracies of 94.66%, 91.16%, 81.01%, and 88.17% for the best variant per dataset. The paper also includes ablations on pretraining duration, masking strategy, input representation, decoder architecture, and backbone freezing, and it releases code and checkpoints.

Significance. If the two-stage cascading design is genuinely responsible for the reported results, the paper provides useful evidence that a vanilla temporal transformer, combined with masked pretraining and a second task-specific transformer, can match or approach graph-based models on standard skeleton benchmarks without explicit spatial graphs. The empirical study is broad, covering three datasets with different modalities, scales, and evaluation protocols, and the public release of code and checkpoints is a concrete reproducibility contribution. The loss functions and attention equations are standard and consistently written, and the reported numbers are externally measured benchmark accuracies, so there is no circularity in the central accuracy claims. However, the paper's stated core novelty—the cascading T2 stage—is never isolated by a control experiment, and the comparison protocol on Penn Action is not shown to be aligned with the cited baselines. These issues directly affect whether the paper's central claims are supported.

major comments (3)
  1. [Cascading Finetuning Component / Insights] The paper's central novelty is the cascading fine-tuning stage (T2 plus cross-attention fusion), and the Insights paragraph asserts that 'the effectiveness of the cascading fine-tuning strategy is consistent across model variants and datasets.' However, no experiment compares the full CascadeFormer pipeline against the standard alternative of fine-tuning the identical pretrained T1 with a linear classification head only. The ablations in Tables 7–8 and supplementary Tables 9–11 vary pretraining duration, masking strategy, input representation, decoder, and backbone freezing, but none isolate T2. Without this control, the reported accuracies in Table 1 could be entirely attributable to masked pretraining of T1 plus a linear head, making the cascade superfluous. This is load-bearing because if T2 provides no measurable gain, the paper's title and central claim collapse even though the accu
  2. [Data Preprocessing / Table 3] The Penn Action evaluation is reported as beating HDM-BG (93.4%) and AOG (85.5%), but the paper states in Data Preprocessing that occluded skeletons were removed using visibility flags. The cited baselines are not shown to use the same filtered protocol. If the removal changes the test set, the comparison in Table 3 is not apples-to-apples, and the 94.66% result may reflect a different (possibly easier) benchmark rather than a model advantage. The authors should either evaluate the baselines under the same filtered protocol, report accuracy on the standard unfiltered test split, or explicitly show that the filtering does not affect the ranking. Without this, the Penn Action comparison is not a controlled claim.
  3. [Performance Comparison / Tables 5-6] The text states that 'even without spatial graphs, our model competes with state-of-the-art graph convolutional approaches,' but the NTU RGB+D 60 comparison set contains only ST-LSTM (2016) and ST-GCN (2018). Modern graph-based methods such as CTR-GCN (cited in the paper's references but not compared) and many later GCNs achieve substantially higher accuracies on these splits. The claim of competing with state-of-the-art graph convolutional approaches is therefore unsupported by the presented comparison. Either the comparison table should be updated to include current GCN and transformer baselines, or the claim should be tempered to 'comparable to ST-GCN,' which is what the data actually show.
minor comments (5)
  1. [Equation (1)] The MSE loss is written as ||masked X - masked X'||^2 over the full tensors, although the text says the loss is computed only on masked joints. A masking operator or index set should be introduced to make the objective precise.
  2. [Training Setup] The architecture hyperparameters for T1 and T2 are not given: embedding dimension, number of layers, number of heads, and decoder dimension are all missing. The claim of reproducibility would be stronger if these were specified in the paper, even if code is available.
  3. [Data Preprocessing (N-UCLA)] The text says the dataset is 'virtually repeat[ed] multiple times' following SkateFormer, but the repetition multiplier is not stated. This is a deterministic preprocessing detail that should be reported.
  4. [Ablation Highlights / Table 7] The number of fine-tuning epochs is fixed at 100 for the pretraining-duration ablation, but the main training setup says the number of epochs varies across variants and datasets. The fine-tuning epoch count for each reported configuration should be stated in the table or text to allow replication.
  5. [Notation in Cross-Attention] In Equation (5), E_cross = Attention(E_pretrain, E_finetune, E_finetune) uses three arguments but the convention for Attention(Q,K,V) is not explicitly repeated. Clarify which argument is Q, K, and V, and whether E_pretrain here is the output of T1 before or after T2 processing.

Circularity Check

0 steps flagged

No circularity: accuracy results are external benchmark measurements, and the missing T2-only control is a support gap rather than a definitional reduction.

full rationale

The paper's central results are accuracies measured on external benchmarks (Penn Action, N-UCLA, NTU RGB+D 60), and no equation in the paper defines those results in terms of a fitted constant or normalizing trick. The masked-pretraining loss is MSE over masked joints and the fine-tuning loss is cross-entropy; these are standard objectives, not self-referential predictions. The paper's claim that the cascading fine-tuning strategy is 'effective' is an interpretation of the full-pipeline accuracy, not a quantity that equals its input by construction. The missing control experiment (linear head on the same pretrained backbone, without T2) is a genuine support gap and a limitation of the evidence, but it is not circularity: the full-pipeline accuracy is not logically forced by anything in the paper, and it could in principle have been worse than a linear-head baseline. The citation to OmniVec2 for hierarchical adaptation is inspirational and is not a self-citation; the reference list contains no overlapping authors with the present paper. The Penn Action occluded-skeleton preprocessing is a protocol concern for comparability, but again not a circular dependency. Therefore no circular step is present, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The ledger is dominated by training hyperparameters typical of an empirical ML paper rather than scientific free parameters: masking ratio, pretraining duration, learning rates, and frame sampling are hand-chosen or lightly ablated. The two genuinely load-bearing assumptions are that Penn Action results stay comparable to baselines after removing occluded skeletons, and that the T2 cascade improves on a linear head, which is asserted but never tested. No new theoretical entities are introduced; T2 is a standard architectural module, not an invented object, so the invented-entities ledger is empty.

free parameters (7)
  • masking_ratio = 0.30 (30% of joints masked, applied across all frames)
    Set by hand; the paper ablates masking strategy (joint vs frame vs none) but not the ratio itself.
  • pretraining_epochs = 100 epochs (default)
    Ablated 1/100/200 on NTU60 CS (Table 7); 100 chosen as the knee of the curve.
  • fine-tuning learning rate = 1e-5 (Penn/N-UCLA), 3e-5 (NTU60 CS), 1e-4 (NTU60 CV)
    Set per dataset 'to ensure optimal performance'; no sensitivity analysis reported.
  • pretraining learning rate = 1e-4 for all datasets
    Fixed value; no sensitivity analysis reported.
  • sampled frame length = 64 frames (N-UCLA and NTU60); per-batch padding on Penn Action
    Chosen to cap padding cost; not ablated.
  • N-UCLA virtual repetition multiplier = not stated in text
    Data multiplication copied from SkateFormer preprocessing; the multiplier value is omitted.
  • architecture hyperparameters (embed dim, depth, heads, decoder size) = not stated in text
    Dimensions are deferred to the released code, so the text does not specify the chosen values.
axioms (6)
  • standard math Standard transformer attention, cross-attention, and MSE pretraining objectives behave as specified (Vaswani et al. 2023, He et al. 2021).
    Invoked throughout the CascadeFormer sections; no formal verification is provided.
  • domain assumption Raw joint coordinates, through the paper's three feature encoders, carry enough information to classify the target actions.
    Grounded in the cited skeleton-recognition literature (Yan et al. 2018, Zhou et al. 2023).
  • domain assumption Masked reconstruction pretraining transfers to the downstream classification task at the finetuning stage.
    Supported by the Table 7 duration ablation, but only on one dataset and one split.
  • domain assumption Keeping only the most active person per NTU60 sequence preserves the label-relevant information.
    Follows the NTU authors' recommendation (Shahroudy et al. 2016), per the Multi-Person Action Analysis section.
  • domain assumption Penn Action accuracies remain comparable to AOG and HDM-BG after removing occluded skeletons.
    Data Preprocessing section; the cited baselines are not shown to use the same protocol.
  • ad hoc to paper The cascading T2 transformer improves classification over a plain linear head on the pretrained T1 embeddings.
    Asserted in the Insights paragraph but never isolated by an ablation in Tables 1-11.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CascadeFormer: A Family of Two-stage Cascading Transformers for Skeleton-based Human Action Recognition." pith.science (2026). https://pith.science/paper/WFLSQ2VL

@misc{pith2026250900692,
  author       = {Pith},
  title        = {Pith review of: CascadeFormer: A Family of Two-stage Cascading Transformers for Skeleton-based Human Action Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WFLSQ2VL}},
  note         = {Machine review of arXiv:2509.00692}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Skeleton-based human action recognition leverages sequences of human joint coordinates to identify actions performed in videos. Owing to the intrinsic spatiotemporal structure of skeleton data, Graph Convolutional Networks (GCNs) have been the dominant architecture in this field. However, recent advances in transformer models and masked pretraining frameworks open new avenues for representation learning. In this work, we propose CascadeFormer, a family of two-stage cascading transformers for skeleton-based human action recognition. Our framework consists of a masked pretraining stage to learn generalizable skeleton representations, followed by a cascading fine-tuning stage tailored for discriminative action classification. We evaluate CascadeFormer across three benchmark datasets (Penn Action N-UCLA, and NTU RGB+D 60), achieving competitive performance on all tasks. To promote reproducibility, we release our code and model checkpoints.

Figures

Figures reproduced from arXiv: 2509.00692 by Alper Yilmaz, Yusen Peng.

Figure 1
Figure 1. Figure 1: Overview of the masked pretraining component in CascadeFormer. A fixed percentage of joints are randomly masked [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the cascading finetuning component in CascadeFormer. The frame embeddings produced by the pre [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Figure 1: The output of T1, denoted as Epretrain, maintains the same shape: Epretrain ∈ R B×T ×embed dim To reconstruct missing joint information, a lightweight lin￾ear decoder is applied to the output of T1. This decoder is used exclusively during pretraining and is discarded dur￾ing downstream fine-tuning. The model is trained to min￾imize the reconstruction error only in the masked joints. Let masked X ∈ R B×C×T … view at source ↗
Figure 3
Figure 3. Figure 3: Feature extraction module in CascadeFormer. All [PITH_FULL_IMAGE:figures/full_fig_p004_3.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

35 extracted references · 16 canonical work pages · 7 internal anchors

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Bertasius, G.; Wang, H.; and Torresani, L. 2021. Is Space-Time Attention All You Need for Video Understanding? arXiv:2102.05095

    Pith/arXiv arXiv 2021

  4. [4]

    Cai, D.; Kang, Y.; Yao, A.; and Chen, Y. 2023. Ske2Grid: Skeleton-to-Grid Representation Learning for Action Recognition. arXiv:2308.07571

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  5. [5]

    Chen, Y.; Zhang, Z.; Yuan, C.; Li, B.; Deng, Y.; and Hu, W. 2021. Channel-wise Topology Refinement Graph Convolution for Skeleton-Based Action Recognition. arXiv:2107.12213

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  6. [6]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv:1810.04805

    Pith/arXiv arXiv 2019

  7. [7]

    Do, J.; and Kim, M. 2024. SkateFormer: Skeletal-Temporal Transformer for Human Action Recognition. arXiv:2403.09508

    Pith/arXiv arXiv 2024

  8. [8]

    Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; Uszkoreit, J.; and Houlsby, N. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv:2010.11929

    Pith/arXiv arXiv 2021

  9. [9]

    Duan, H.; Zhao, Y.; Chen, K.; Lin, D.; and Dai, B. 2022. Revisiting Skeleton-based Action Recognition. arXiv:2104.13586

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  10. [10]

    He, K.; Chen, X.; Xie, S.; Li, Y.; Dollár, P.; and Girshick, R. 2021. Masked Autoencoders Are Scalable Vision Learners. arXiv:2111.06377

    Pith/arXiv arXiv 2021

  11. [11]

    Kim, S.; Ahn, D.; and Ko, B. C. 2023. Cross-Modal Learning with 3D Deformable Attention for Action Recognition. arXiv:2212.05638

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  12. [12]

    Lee, I.; Kim, D.; Kang, S.; and Lee, S. 2017. Ensemble Deep Learning for Skeleton-Based Action Recognition Using Temporal Sliding LSTM Networks. In 2017 IEEE International Conference on Computer Vision (ICCV), 1012--1020

    work page 2017

  13. [13]

    Liu, H.; Li, C.; Wu, Q.; and Lee, Y. J. 2023. Visual Instruction Tuning. arXiv:2304.08485

    Pith/arXiv arXiv 2023

  14. [14]

    Liu, J.; Shahroudy, A.; Xu, D.; and Wang, G. 2016. Spatio-Temporal LSTM with Trust Gates for 3D Human Action Recognition. arXiv:1607.07043

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  15. [15]

    Liu, M.; Liu, H.; and Chen, C. 2017. Enhanced Skeleton Visualization for View Invariant Human Action Recognition. Pattern Recognition, 68: 346--362

    work page 2017

  16. [16]

    Loshchilov, I.; and Hutter, F. 2017. SGDR: Stochastic Gradient Descent with Warm Restarts. arXiv:1608.03983

    Pith/arXiv arXiv 2017

  17. [17]

    Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. arXiv:1711.05101

    Pith/arXiv arXiv 2019

  18. [18]

    X.; Xiong, C.; and Zhu, S.-C

    Nie, B. X.; Xiong, C.; and Zhu, S.-C. 2015. Joint action recognition and pose estimation from video. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 1293--1301

    work page 2015

  19. [19]

    Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; Desmaison, A.; Kopf, A.; Yang, E.; DeVito, Z.; Raison, M.; Tejani, A.; Chilamkurthy, S.; Steiner, B.; Fang, L.; Bai, J.; and Chintala, S. 2019. PyTorch: An Imperative Style, High-Performance Deep Learning Library. In Advances in Neural ...

    work page 2019

  20. [20]

    W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I

    Radford, A.; Kim, J. W.; Hallacy, C.; Ramesh, A.; Goh, G.; Agarwal, S.; Sastry, G.; Askell, A.; Mishkin, P.; Clark, J.; Krueger, G.; and Sutskever, I. 2021. Learning Transferable Visual Models From Natural Language Supervision. arXiv:2103.00020

    Pith/arXiv arXiv 2021

  21. [21]

    Rao, Y.; Zhao, W.; Liu, B.; Lu, J.; Zhou, J.; and Hsieh, C.-J. 2021. DynamicViT: Efficient Vision Transformers with Dynamic Token Sparsification. arXiv:2106.02034

    Pith/arXiv arXiv 2021

  22. [22]

    Ridnik, T.; Ben-Baruch, E.; Noy, A.; and Zelnik-Manor, L. 2021. ImageNet-21K Pretraining for the Masses. arXiv:2104.10972

    Pith/arXiv arXiv 2021

  23. [23]

    Ruder, S. 2017. An overview of gradient descent optimization algorithms. arXiv:1609.04747

    Pith/arXiv arXiv 2017

  24. [24]

    Shahroudy, A.; Liu, J.; Ng, T.-T.; and Wang, G. 2016. NTU RGB+D: A Large Scale Dataset for 3D Human Activity Analysis. arXiv:1604.02808

    Pith/arXiv arXiv 2016

  25. [25]

    Shi, L.; Zhang, Y.; Cheng, J.; and Lu, H. 2019. Skeleton-Based Action Recognition with Directed Graph Neural Networks . In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 7912--7921. Long Beach, CA

    work page 2019

  26. [26]

    Srivastava, S.; and Sharma, G. 2023. OmniVec: Learning robust representations with cross modal sharing. arXiv:2311.05709

    Pith/arXiv arXiv 2023

  27. [27]

    Srivastava, S.; and Sharma, G. 2024. OmniVec2 - A Novel Transformer Based Network for Large Scale Multimodal and Multitask Learning. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 27402--27414

    work page 2024

  28. [28]

    N.; Kaiser, L.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2023. Attention Is All You Need. arXiv:1706.03762

    Pith/arXiv arXiv 2023

  29. [29]

    wang, J.; Nie, X.; Xia, Y.; Wu, Y.; and Zhu, S.-C. 2014. Cross-view Action Modeling, Learning and Recognition. arXiv:1405.2941

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  30. [30]

    Wang, Q.; Shi, S.; He, J.; Peng, J.; Liu, T.; and Weng, R. 2023. IIP-Transformer: Intra-Inter-Part Transformer for Skeleton-Based Action Recognition. In 2023 IEEE International Conference on Big Data (BigData), 936–945. IEEE

    work page 2023

  31. [31]

    Yan, S.; Xiong, Y.; and Lin, D. 2018. Spatial Temporal Graph Convolutional Networks for Skeleton-Based Action Recognition. arXiv:1801.07455

    Pith/arXiv arXiv 2018

  32. [32]

    Zhang, C.; Gupta, A.; and Zisserman, A. 2021. Temporal Query Networks for Fine-grained Video Understanding. arXiv:2104.09496

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  33. [33]

    Zhang, W.; Zhu, M.; and Derpanis, K. G. 2013. From Actemes to Action: A Strongly-Supervised Representation for Detailed Action Understanding. In 2013 IEEE International Conference on Computer Vision, 2248--2255

    work page 2013

  34. [34]

    Zhao, R.; Xu, W.; Su, H.; and Ji, Q. 2019. Bayesian Hierarchical Dynamic Model for Human Action Recognition. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 7725--7734

    work page 2019

  35. [35]

    Zhou, Y.; Cheng, Z.-Q.; Li, C.; Fang, Y.; Geng, Y.; Xie, X.; and Keuper, M. 2023. Hypergraph Transformer for Skeleton-based Action Recognition. arXiv:2211.09590

    Pith/arXiv arXiv 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.