Pith. sign in

REVIEW 4 major objections 5 minor 40 references

With 10 samples per class, a text-to-motion diffusion prior can be distilled into a specialized action generator, raising median top-1 HAR accuracy from 63.1% to 86.2%.

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 16:46 UTC pith:JW3B6DX7

load-bearing objection The CLIP-based kinetic-mining pipeline is a genuinely new way to turn a T2M prior into few-shot HAR augmentation, but the paper's own ablation undercuts its headline configuration and the gain over a plain pretrained MDM is only 2.3 pp. the 4 major comments →

arxiv 2512.11654 v3 pith:JW3B6DX7 submitted 2025-12-12 cs.CV

Kinetic Mining in Context: Few-Shot Action Synthesis via Text-to-Motion Distillation

Luca Cazzola , Ahed Alboody This is my paper
classification cs.CV
keywords few-shot action synthesistext-to-motionhuman activity recognitionskeleton-based HARdiffusion modelsdata augmentationkinetic miningteacher-student distillation
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.

KineMIC is a teacher-student transfer framework that converts a generalist text-to-motion (T2M) diffusion model, trained on long descriptive captions, into a few-shot Action-to-Motion generator for skeleton-based human activity recognition (HAR). The paper claims that semantic correspondences in a shared text-embedding space (CLIP) provide sufficiently reliable soft supervision to locate kinematically relevant segments from a large unlabeled source motion collection, and that distilling those segments into a lightly fine-tuned student yields synthetic motions that improve a downstream classifier. On three NTU RGB+D 120 actions with 10 real samples per class, KineMIC achieves 86.2% median top-1 accuracy, a +23.1 percentage-point gain over real-data-only training (63.1%) and near parity with a prior GAN-based augmentation method (86.4%). The authors' own ablation shows the dynamic window weighting, not the teacher-distillation term, is the component that secures the gain; the central risk is that semantic similarity does not always track kinematic relevance.

Core claim

The central discovery is that a generalist diffusion motion prior, whose training objective favors broad artistic motion, can be repurposed for HAR's kinematically precise action classes by mining context. The method pairs each scarce target sample with semantically matched source captions via CLIP text embeddings, then learns frame-level correspondences through a contrastively trained module whose attention identifies the source window most kinematically consistent with the target action. That window is used as a pseudo-labeled training signal, with a similarity-based weight modulating its influence. The result is that the student remains specialized without collapsing: with 30 real samples

What carries the argument

The load-bearing component is the Mining In Context (MIC) module, an attention-enhanced bidirectional GRU that ingests frame-wise tokens from both the frozen teacher and the trainable student streams and produces context-aware latent vectors. A Soft Nearest Neighbors contrastive loss aligns latents of soft-positive pairs (target sample vs. retrieved source motion), so the attention distribution learns which frames of the long source motion are kinematically relevant. A max-cumulative-attention window extraction selects the matching subsequence, which is then used as a pseudo-label for the student's reconstruction and distillation losses, scaled by a dynamic window weight (dww) equal to the c

Load-bearing premise

The whole retrieval-and-mining chain assumes that captions whose text embeddings are similar to a target label describe motions whose kinematics are actually close to the target motion; when that fails (e.g., 'punch' vs. 'kick'), the mined windows are kinematically wrong and the filtering weight cannot recover.

What would settle it

Run KineMIC with the same hyperparameters on action classes that share high CLIP text similarity but different kinematics, such as punch vs. kick or wave vs. salute, using 10 samples per class. If median top-1 accuracy with augmentation drops below the real-data-only baseline (or shows no improvement), the semantic-proxy assumption fails; conversely, if dww filters bad windows successfully, the approach generalizes beyond the three easy actions tested.

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

If this is right

  • If T2M priors can be specialized this cheaply, few-shot HAR augmentation no longer depends on training a GAN from scratch; a frozen generalist plus lightweight adaptation suffices.
  • The +23.1 percentage-point gain over real-data-only training indicates that synthetic augmentation from a semantic-matched prior is viable even with only 10 samples per class.
  • The framework's components transfer to any pair of source text-motion collection and target action-label domain sharing a text-embedding space, provided the mining filter (dww) is retained.
  • The ablation that adding distillation alone hurts (83.78% vs. 85.21%) implies the teacher's generalist behavior must be filtered, not blindly imitated.
  • KineMIC's dual conditioning enables a form of motion composition between action id and text prompt, as in the 'stretch while jumping' example.

Where Pith is reading between the lines

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

  • Editorial: The decisive test is whether the semantic-proxy assumption holds on hard pairs (punch vs. kick, or actions with near-duplicate text embeddings but different kinematics). Running the same protocol on such classes would likely reveal that dww alone cannot recover when no retrieved window is kinematically correct.
  • Editorial: The method's scalability to larger label sets is untested. With more action classes, the top-k retrieval per class could overlap heavily, and the contrastive batch may need harder negatives; a natural extension is to increase k or add class-aware negative sampling.
  • Editorial: Because the pipeline already yields composition between action and text, one could test whether KineMIC can generate motions for unseen compositions of seen actions without retraining, which would extend its use beyond augmentation to interactive motion editing.

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

4 major / 5 minor

Summary. The paper proposes KineMIC, a teacher-student framework that adapts a pre-trained text-to-motion (T2M) diffusion model (MDM) to a few-shot skeleton-based HAR task. The target action labels are matched to T2M source captions via CLIP text embeddings (soft positive search); a MIC module (attention-enhanced biGRU) aligns target and source motion latents with a contrastive loss; relevant source-motion windows are mined by attention and used as pseudo-training samples for the student, with a dynamic window weighting (dww) and a window-distillation loss. Experiments on three NTU RGB+D 120 actions with 10 samples per class report 86.2% median top-1 accuracy versus 63.1% for real-data-only training, approaching the 86.4% of prior GAN-based work. The paper includes ablations, diversity/multimodality metrics, and qualitative examples.

Significance. If the results hold, the paper demonstrates a meaningful new use of general T2M diffusion priors for few-shot HAR augmentation, offering an alternative to GAN-based data synthesis. The method is relatively simple, uses standard components, and the experimental setup is mostly described in sufficient detail to reproduce. The honest ablation study is a strength: it identifies the dynamic window weighting as the key component and reports that the distillation loss is harmful. However, the paper's central claim is weakened by a load-bearing inconsistency between the described complete loss (Eq. 7, which includes L_dist) and the ablation result that the best configuration excludes L_dist. The reported +23.1 pp gain is also measured against a weak real-data-only baseline; the improvement over simply using the pre-trained MDM as a generator is only about +2.3 pp. The paper's own stated limitation that semantic similarity is an unreliable proxy for kinematic relevance is not tested on hard cases, so the generality of the framework for arbitrary actions is unsubstantiated. The contribution is promising but needs revision to align claims with evidence.

major comments (4)
  1. [§5.2 Table 1 vs §5.3 Table 2 and Eq. (7)] The headline 'KineMIC 86.2' in Table 1 is not tied to a stated configuration. Table 2 shows that adding L_dist (Eq. 5) to the base model reduces accuracy from 85.21±2.24 to 83.78±2.38, and that the full Eq. (7) loss (L_dist + dww) gives 84.94±2.37, below dww-only at 86.41±0.95. If the 86.2% result comes from the dww-only configuration, then the paper's described 'complete multi-objective loss' and the title's 'distillation' mechanism are not what drive the gain; if it comes from the full Eq. (7) loss, it contradicts Table 2. Please state explicitly which loss configuration produced each row of Table 1 and revise the abstract/conclusions so that the described method matches the best evaluated configuration.
  2. [§4.5 and §5.3] The window-distillation loss L_dist is presented as a core component of the framework (Eq. 5 and Eq. 7), but the ablation shows it is net harmful for accuracy while increasing diversity/multimodality. The paper acknowledges this in the Table 2 discussion, but the method description, abstract, and contributions still frame distillation as central. The paper should be restructured so that the actual operating method is the best configuration (base + dynamic window weighting), with L_dist explicitly described as an optional/ablated variant. Otherwise the scientific claim is not internally consistent.
  3. [§6 and §4.2] The core assumption that CLIP text-space similarity is a proxy for kinematic relevance is acknowledged in the limitations as unreliable for pairs like punch vs. kick, yet all three evaluated actions (running on spot, side kick, stretch on self) are simple actions that the paper itself says a generalist T2M model can already generate. This means the evaluation does not test the framework under the failure mode the authors identify. To support the general claim of adapting T2M priors to arbitrary HAR actions, the paper should either include harder, semantically confusable action classes or explicitly restrict the claim to actions for which semantic retrieval is reliable.
  4. [§5.3, Table 2] The label 'dww only' is ambiguous. The dww score (Eq. 6) is computed from MIC latents z_T and z_P*, and the 'Base' row already includes the contrastive loss (Eq. 1) and target reconstruction (Eq. 3). If 'dww only' means 'Base + dww', this should be stated explicitly; as written, a reader could infer that the dynamic weighting works without the contrastive alignment, which would be circular because dww is a function of the aligned MIC embeddings.
minor comments (5)
  1. [Abstract] The phrase '+23.1% accuracy points improvement' should be '+23.1 percentage points' or '+23.1 pp' to avoid ambiguity with relative improvement.
  2. [Table 1 caption] Typo: 'T able' should be 'Table'.
  3. [Fig. 4 caption] Action labels are inconsistent: figure says 'Run on spot', 'Side kick', 'Stretch on self' while the text uses 'running on spot', 'side kick', 'stretch on self'. Please standardize.
  4. [§5.1] The notation D¯T is introduced but used inconsistently (e.g., '¯T ⊂ T' vs. 'D¯T'). Please use a single notation for the few-shot training subset.
  5. [§5.2] Table 1 reports median top-1 accuracy while Table 2 reports mean±std. The text should state this explicitly near the tables so readers do not compare the two tables directly as if they are the same statistic.

Circularity Check

1 steps flagged

No central derivation-equals-input circularity; the dww window-quality filter is self-referential because it is computed from the same MIC similarity that Eq. 1 was trained to maximize for soft-positive pairs.

specific steps
  1. fitted input called prediction [Sec. 4.3 Eq. (1) vs Sec. 4.5 Eq. (6); Sec. 4.4 window mining]
    "MIC is trained to align z¯T_i with z¯P_j from soft-positive pairs sharing the same target class y. ... we compute a window quality score: dww = (1 + sim(z¯T, z^P*))/2 ... z^P* is computed with gradients detached, ensuring MIC remains trained exclusively by Eq. 1. The score dww scales both reconstruction (Eq. 4) and distillation (Eq. 5) losses, giving higher weight to more reliable prior windows."

    The 'quality' score is the cosine similarity that Lcon (Eq. 1) was trained to maximize between target latents and soft-positive source latents. Soft positives are retrieved by CLIP semantics and can be kinematically wrong, as the paper concedes for 'punch' vs 'kick'; MIC's attention is trained to pick frames minimizing this alignment loss, so the mined-window latent z^P* inherits the same trained similarity. Thus dww is high by construction for any retrieved soft-pair, rather than measuring independent kinematic reliability. Using it to up/down-weight windows renames the contrastive training objective as a filter.

full rationale

No self-citations, imported uniqueness theorems, or ansatz-by-citation were found. The headline accuracy gain (+23.1 pp) is evaluated on held-out real NTU RGB+D 120 test samples, so the central performance claim is externally grounded and does not reduce to the training objectives. The paper's own ablations show the full Eq. 7 configuration underperforms dww-only (84.94±2.37 vs 86.41±0.95), and the configuration behind Table 1's 86.2 is not stated; that is a reproducibility/correctness concern, not a circularity. The only construction-level circularity is the dww score: it uses MIC embeddings that Eq. 1 explicitly trains to make soft-positive pairs similar, and the mining attention is trained toward the same objective, so Eq. 6 is the training objective renamed as a quality filter. This affects a supporting component of the framework, not the external benchmark, so the overall circularity score is moderate.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 2 invented entities

The framework is empirical; it imports CLIP and MDM from prior work and postulates that semantic similarity in CLIP space tracks kinematic similarity, that the MDM prior transfers to VIBE-extracted NTU skeletons, and that a frozen teacher guides LoRA fine-tuning without collapse. These are domain assumptions, not proven.

free parameters (6)
  • soft positive retrieval count k = 250
    Chosen 'to effectively balance diversity and semantic relevance' (§5.1); not ablated.
  • contrastive temperature τ = 0.07
    Standard CLIP temperature, fixed in Eq. 1; not ablated.
  • LoRA rank/α/dropout = 16 / 32 / 0.1
    Standard LoRA config from prior work; no ablation on HAR transfer.
  • loss weights λ_T_rec, λ_con, λ_P*_rec, λ_dist = 1.0 each
    All set to 1.0 without tuning; L_dist's contribution is negative in Table 2, suggesting weights matter.
  • CFG guidance scale = 2.5
    Inherited from MDM; no ablation.
  • training steps = 5000
    Chosen with AdamW LR 2e-5; no early stopping analysis.
axioms (4)
  • domain assumption CLIP text embedding similarity between HAR label prompts and T2M captions is a reliable proxy for kinematic relevance (Sec. 4.2).
    Core retrieval-and-mining premise; authors themselves flag punch-vs-kick failure in Sec. 6.
  • domain assumption MDM pretrained on HumanML3D transfers to NTU skeletons after VIBE estimation, joint dropping, fps downsampling, and HumanML3D normalization (Sec. 5.1).
    No quantitative check that VIBE-extracted skeleton statistics match MoCap distribution.
  • domain assumption A frozen teacher GP guides LoRA fine-tuned student GT without collapse, preserving prior kinematics (Sec. 4.1).
    Empirical claim; ablation shows naive LoRA collapses (diversity 9.45), so the mechanism is delicate.
  • standard math Diffusion model forward/backward processes and contrastive (Soft Nearest Neighbor) losses are valid background (Eq. 1-5).
    Standard formulations from cited MDM and SNN literature.
invented entities (2)
  • MIC module (attention-enhanced biGRU encoder) no independent evidence
    purpose: Aggregates frame-wise teacher/student tokens into context-aware latents zP, zT for contrastive alignment and window mining (Eq. 1-2).
    Novel architecture with no formal guarantees or external benchmark; validated only within this paper's pipeline.
  • Learnable action embedding no independent evidence
    purpose: Conditioning vector for discrete HAR labels in the student stream (Fig. 2).
    New trainable parameter; no independent justification beyond empirical gain.

pith-pipeline@v1.3.0-alltime-deepseek · 10217 in / 12645 out tokens · 108388 ms · 2026-08-03T16:46:55.837608+00:00 · methodology

0 comments
read the original abstract

The acquisition cost for large, annotated motion datasets remains a critical bottleneck for skeletal-based Human Activity Recognition (HAR). Although Text-to-Motion (T2M) generative models offer a compelling, scalable source of synthetic data, their training objectives, which emphasize general artistic motion, and dataset structures fundamentally differ from HAR's requirements for kinematically precise, class-discriminative actions. This disparity creates a significant domain gap, making generalist T2M models ill-equipped for generating motions suitable for HAR classifiers. To address this challenge, we propose KineMIC (Kinetic Mining In Context), a transfer learning framework for few-shot action synthesis. KineMIC adapts a T2M diffusion model to an HAR domain by hypothesizing that semantic correspondences in the text encoding space can provide soft supervision for kinematic distillation. We operationalize this via a kinetic mining strategy that leverages CLIP text embeddings to establish correspondences between sparse HAR labels and T2M source data. This process guides fine-tuning, transforming the generalist T2M backbone into a specialized few-shot Action-to-Motion generator. We validate KineMIC using HumanML3D as the source T2M dataset and a subset of NTU RGB+D 120 as the target HAR domain, randomly selecting just 10 samples per action class. Our approach generates significantly more coherent motions, providing a robust data augmentation source that delivers a +23.1% accuracy points improvement. Animated illustrations and supplementary materials are available at https://lucazzola.github.io/kinemic-page/.

Figures

Figures reproduced from arXiv: 2512.11654 by Ahed Alboody, Luca Cazzola.

Figure 1
Figure 1. Figure 1: Kinetic Mining in Context. The target action sample (bottom) is used to contextualize the search within the large source data sample (top), establishing soft pairs. The mining operation identifies a kinematically relevant segment (in orange) from the source data. performance of deep learning models for this task is fundamentally limited by the availability of large-scale, accurately annotated datasets. The… view at source ↗
Figure 2
Figure 2. Figure 2: Source stream (red): c P from CLIP text embeddings + timestep. Target stream (blue): c T from action-label text embedding + learnable action embedding + timestep. Dotted lines denote weight sharing. 1. Conditioning modalities differ fundamentally (semantic gap). The source domain uses high-variance, free-form text, while the target domain employs discrete action labels. 2. There is a discrepancy in the mot… view at source ↗
Figure 3
Figure 3. Figure 3: (Left) Core pipeline: dashed lines show detached gradients; dotted lines show weight sharing. MIC receives frame-wise tokens from teacher (G P ) and student (G T ) streams, producing latents z P and z T¯ for contrastive alignment and window mining. (Right) MIC module: attention-enhanced biGRU [1,4] encoder aggregating frame-wise motion tokens into context-aware latents. the student inherits the frozen text… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Comparison of Generated Motions. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Synthetic data scalability. When the ST-GCN classifier is trained using only real D T data, training exhibits high variance and poor performances. When synthetic data from KineMIC is concatenated, accuracy is consistently much higher. 5.3 Ablation Study In [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Motion composition via dual conditioning. [PITH_FULL_IMAGE:figures/full_fig_p013_6.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

40 extracted references · 5 linked inside Pith

  1. [1]

    In: ICLR (2015)

    Bahdanau, D., Cho, K., Bengio, Y.: Neural machine translation by jointly learning to align and translate. In: ICLR (2015)

    2015

  2. [2]

    In: ICCV (2021)

    Chen, Y., Zhang, Z., Yuan, C., Li, B., Deng, Y., Hu, W.: Channel-wise topol- ogy refinement graph convolution for skeleton-based action recognition. In: ICCV (2021)

    2021

  3. [3]

    In: CVPR (2022)

    Chi, H.G., Ha, M.H., Chi, S., Lee, S.W., Huang, Q., Ramani, K.: InfoGCN: Repre- sentation learning for human skeleton-based action recognition. In: CVPR (2022)

    2022

  4. [4]

    arXiv:1412.3555 (2014)

    Chung, J., Gulcehre, C., Cho, K., Bengio, Y.: Empirical evaluation of gated recur- rent neural networks on sequence modeling. arXiv:1412.3555 (2014)

    Pith/arXiv arXiv 2014

  5. [5]

    In: WACV (2022)

    Degardin, B., Neves, J., Lopes, V., Brito, J., Yaghoubi, E., Proenca, H.: Generative adversarial graph convolutional networks for human action synthesis. In: WACV (2022)

    2022

  6. [6]

    Expert Syst

    Dentamaro, V., Gattulli, V., Impedovo, D., Manca, F.: Human activity recognition with smartphone-integrated sensors: A survey. Expert Syst. Appl. (2024)

    2024

  7. [7]

    In: ECCV (2024)

    Do, J., Kim, M.: Skateformer: Skeletal-temporal transformer for human action recognition. In: ECCV (2024)

    2024

  8. [8]

    In: CVPR (2015)

    Du, Y., Wang, W., Wang, L.: Hierarchical recurrent neural network for skeleton based action recognition. In: CVPR (2015)

    2015

  9. [9]

    In: ACM MM (2022)

    Duan,H.,Wang,J.,Chen,K.,Lin,D.:PYSKL:Towardsgoodpracticesforskeleton action recognition. In: ACM MM (2022)

    2022

  10. [10]

    In: CVPR (2022)

    Duan, H., Zhao, Y., Chen, K., Lin, D., Dai, B.: Revisiting skeleton-based action recognition. In: CVPR (2022)

    2022

  11. [11]

    In: ICML (2019)

    Frosst, N., Papernot, N., Hinton, G.: Analyzing and improving representations with the soft nearest neighbor loss. In: ICML (2019)

    2019

  12. [12]

    In: WACV (2024)

    Fukushi, K., Nozaki, Y., Nishihara, K., Nakahara, K.: Few-shot generative model for skeleton-based human action synthesis using cross-domain adversarial learning. In: WACV (2024)

    2024

  13. [13]

    In: CVPR (2024)

    Guo, C., Mu, Y., Javed, M.G., Wang, S., Cheng, L.: Momask: Generative masked modeling of 3d human motions. In: CVPR (2024)

    2024

  14. [14]

    In: CVPR (2022)

    Guo, C., Zou, S., Zuo, X., Wang, S., Ji, W., Li, X., Cheng, L.: Generating diverse and natural 3d human motions from text. In: CVPR (2022)

    2022

  15. [15]

    In: ACM MM (2020)

    Guo, C., Zuo, X., Wang, S., Zou, S., Sun, Q., Deng, A., Gong, M., Cheng, L.: Action2motion: Conditioned generation of 3d human motions. In: ACM MM (2020)

    2020

  16. [16]

    In: ICLR (2022)

    Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W.: LoRA: Low-rank adaptation of large language models. In: ICLR (2022)

    2022

  17. [17]

    Sensors (Basel, Switzerland) (2023)

    Hung-Cuong, N., Nguyen, T.H., Scherer, R., Le, V.H.: Deep learning for human activity recognition on 3d human skeleton: Survey and comparative study. Sensors (Basel, Switzerland) (2023)

    2023

  18. [18]

    In: ECCV (2024) Kinetic Mining in Context 15

    Khirodkar, R., Bagautdinov, T., Martinez, J., Zhaoen, S., James, A., Selednik, P., Anderson, S., Saito, S.: Sapiens: Foundation for human vision models. In: ECCV (2024) Kinetic Mining in Context 15

    2024

  19. [19]

    In: CVPR (2020)

    Kocabas, M., Athanasiou, N., Black, M.J.: VIBE: Video inference for human body pose and shape estimation. In: CVPR (2020)

    2020

  20. [20]

    arXiv:2506.07612 (2025)

    Leng, Z., Iyer, A., Plotz, T.: Scaling human activity recognition: A compar- ative evaluation of synthetic data generation and augmentation techniques. arXiv:2506.07612 (2025)

    Pith/arXiv arXiv 2025

  21. [21]

    IEEE Trans

    Liu, J., Shahroudy, A., Perez, M., Wang, G., Duan, L., Kot, A.C.: NTU RGB+D 120: A large-scale benchmark for 3d human activity understanding. IEEE Trans. Pattern Anal. Mach. Intell. (2020)

    2020

  22. [22]

    In: CVPR (2020)

    Liu, Z., Zhang, H., Chen, Z., Wang, Z., Ouyang, W.: Disentangling and unifying graph convolutions for skeleton-based action recognition. In: CVPR (2020)

    2020

  23. [23]

    Lupion, M., Cruciani, F., Cleland, I., Nugent, C., Ortigosa, P.: Data augmenta- tion for human activity recognition with generative adversarial networks. IEEE J. Biomed. Health Informatics (2024)

    2024

  24. [24]

    In: ICCV (2019)

    Mahmood, N., Ghorbani, N., Troje, N.F., Pons-Moll, G., Black, M.J.: AMASS: Archive of motion capture as surface shapes. In: ICCV (2019)

    2019

  25. [25]

    In: ICCV (2021)

    Petrovich, M., Black, M.J., Varol, G.: Action-conditioned 3d human motion syn- thesis with transformer VAE. In: ICCV (2021)

    2021

  26. [26]

    In: ECCV (2022)

    Petrovich, M., Black, M.J., Varol, G.: TEMOS: Generating diverse human motions from textual descriptions. In: ECCV (2022)

    2022

  27. [27]

    In: ICPR (2020)

    Plizzari, C., Cannici, M., Matteucci, M.: Spatial temporal transformer network for skeleton-based action recognition. In: ICPR (2020)

    2020

  28. [28]

    In: ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  29. [29]

    In: AISTATS (2007)

    Salakhutdinov, R., Hinton, G.: Learning a nonlinear embedding by preserving class neighbourhood structure. In: AISTATS (2007)

    2007

  30. [30]

    arXiv:2503.19557 (2025)

    Sawdayee, H., Guo, C., Tevet, G., Zhou, B., Wang, J., Bermano, A.H.: Dance like a chicken: Low-rank stylization for human motion diffusion. arXiv:2503.19557 (2025)

    Pith/arXiv arXiv 2025

  31. [31]

    In: ICLR (2024)

    Shafir, Y., Tevet, G., Kapon, R., Bermano, A.H.: Human motion diffusion as a generative prior. In: ICLR (2024)

    2024

  32. [32]

    In: CVPR (2016)

    Shahroudy, A., Liu, J., Ng, T.T., Wang, G.: NTU RGB+D: a large scale dataset for 3d human activity analysis. In: CVPR (2016)

    2016

  33. [33]

    In: CVPR (2019)

    Shi, L., Zhang, Y., Cheng, J., Lu, H.: Two-stream adaptive graph convolutional networks for skeleton-based action recognition. In: CVPR (2019)

    2019

  34. [34]

    In: ICLR (2025)

    Tevet, G., Raab, S., Cohan, S., Reda, D., Luo, Z., Peng, X.B., Bermano, A.H., van de Panne, M.: CLoSD: closing the loop between simulation and diffusion for multi-task character control. In: ICLR (2025)

    2025

  35. [35]

    In: ICLR (2023)

    Tevet, G., Raab, S., Gordon, B., Shafir, Y., Cohen-Or, D., Bermano, A.H.: Human motion diffusion model. In: ICLR (2023)

    2023

  36. [36]

    In: CVPR (2025)

    Wang, Y., Sun, Y., Patel, P., Daniilidis, K., Black, M.J., Kocabas, M.: PromptHMR: Promptable human mesh recovery. In: CVPR (2025)

    2025

  37. [37]

    International Journal of Computer Vision (2025)

    Wanyan, Y., Yang, X., Dong, W., Xu, C.: A comprehensive review of few-shot action recognition. International Journal of Computer Vision (2025)

    2025

  38. [38]

    In: AAAI (2018)

    Yan, S., Xiong, Y., Lin, D.: Spatial temporal graph convolutional networks for skeleton-based action recognition. In: AAAI (2018)

    2018

  39. [39]

    arXiv:2503.06955 (2025)

    Zhang, Z., Wang, Y., Mao, W., Li, D., Zhao, R., Wu, B., Song, Z., Zhuang, B., Reid, I., Hartley, R.: Motion anything: Any to motion generation. arXiv:2503.06955 (2025)

    Pith/arXiv arXiv 2025

  40. [40]

    arXiv:2301.03949 (2023)

    Zhao, M., Liu, M., Ren, B., Dai, S., Sebe, N.: Modiff: Action-conditioned 3d motion generation with denoising diffusion probabilistic models. arXiv:2301.03949 (2023)

    Pith/arXiv arXiv 2023