Pith. sign in

REVIEW 3 major objections 5 minor 66 references

Synthetic Human Action Video Data Generation with Pose Transfer

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adding synthetic clips made by pose transfer—reenacting real actions with 3D Gaussian avatars of new identities on varied backgrounds—improves action recognition accuracy on Toyota Smarthome and NTU RGB+D in baseline, one-shot, and…

desk verdict A useful, reproducible synthetic-video pipeline whose main accuracy claims are undercut by a missing matched real-data control. read the letter →

arxiv 2506.09411 v1 pith:QPE24IAO submitted 2025-06-11 cs.CV cs.AI

Vaclav Knapp , Matyas Bohacek This is my paper · ORCID
classification cs.CVcs.AI
keywords syntheticdataposetransferhumanactionrecognition3DGaussianavatarsvideogenerationfew-shotlearningidentitydiversitybackground
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that synthetic human-action videos made by pose transfer can serve as effective training data for action recognition models. The method takes real videos of human actions, reenacts them with 3D Gaussian avatars of novel identities, and composites the animated figures onto varied background images. On 16-class subsets of Toyota Smarthome and NTU RGB+D, adding these synthetic clips raised ResNet accuracy from 20.46% to 51.15% and from 8.66% to 42.98%, and SlowFast from 38.35% to 55.64% and from 26.75% to 36.29%. The paper also reports that one-shot and few-shot models improve as the number of synthetic samples grows, arguing the method can scale few-shot datasets. If correct, this offers a route to generating labeled video data for underrepresented identities, backgrounds, and low-resource action-recognition tasks.

What carries the argument

The load-bearing object is a modified ExAvatar pipeline: an expressive whole-body 3D Gaussian avatar model that is fully animatable. For each novel identity, the method builds an avatar from a short rotation video, extracting SMPL-X meshes, depth, body keypoints, segmentation masks, facial expressions, and hand poses; for each reference action video it extracts the corresponding pose features. Animation reenacts the reference pose sequence with the new identity, first on a white background and then composited onto randomly selected background images. This machinery is what gives the method its purported advantage: the 3D avatar guarantees pose control while the novel identities and backgrounds supply the diversity that real datasets often lack.

What would settle it

Retrain the same ResNet on synthetic clips in which each clip's pose sequence has been swapped with the pose sequence of a different action class, keeping identities and backgrounds fixed; if accuracy stays near the real-plus-synthetic level, the gain comes from appearance and scene cues rather than from action transfer.

Watch

Extended reading notes

Core claim

On its own terms, the paper's finding is that synthetic videos produced by reenacting reference action clips with controllable 3D Gaussian avatars of novel identities, then compositing them on varied backgrounds, are effective training data for video action classifiers. In the baseline experiments, ResNet accuracy on a 16-class Toyota Smarthome subset rose from 20.46% to 51.15% when synthetic clips were added, and from 8.66% to 42.98% on a comparable NTU RGB+D subset; SlowFast rose from 38.35% to 55.64% and from 26.75% to 36.29%. One-shot and few-shot experiments show accuracy climbing as the number of synthetic samples per class increases, and the paper concludes the method can scale few-shot datasets by adding identity and background diversity. The paper also contributes RANDOM People, an open dataset of identity videos, avatar models, synthetic clips, and background images.

Load-bearing premise

The load-bearing premise is that pose transfer preserves the action class well enough for training on synthetic clips to transfer to real clips; the paper verifies this only qualitatively, not with a quantitative action-fidelity measure.

Editorial extensions

If this is right

  • Action classifiers trained with these synthetic clips need substantially less real labeled video to reach usable accuracy, which is the paper's central quantitative result.
  • The same recipe can be applied to sign language translation, gesture recognition, and autonomous driving perception, where labeled video is scarce.
  • Adding avatars of underrepresented demographic groups and varied backgrounds can reduce identity and scene bias in video models.
  • Scaling beyond the 15 identities used in the experiments is a direct path to further gains, given the observed upward trends.
  • The method's current limits—single protagonists, no object interaction, scene plausibility—are ceilings inherited from the backbone, not intrinsic to the data-generation idea.

Reading between the lines

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

  • We infer that part of the accuracy gain may come from the added identity and background diversity rather than from faithful action transfer; the paper does not ablate this, so the size of each contribution is unknown.
  • We infer that the method's ceiling tracks the pose-transfer backbone: if a future backbone adds object interaction and multi-person support, the same pipeline extends to cooking, tool use, and social actions.
  • We infer the synthetic data could support fairness evaluations by rebalancing identity distributions, but that requires per-subgroup accuracy reporting, which this paper does not include.
  • A direct testable extension is to train on synthetic clips where the pose sequence is intentionally mismatched to the label; matching accuracy would signal that the model learns appearance rather than action.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a pipeline for synthetic human action video generation based on pose transfer, using ExAvatar 3D Gaussian avatars to reenact actions from reference videos with novel identities and backgrounds. The authors evaluate the method on 16-class subsets of Toyota Smarthome and NTU RGB+D, reporting large classification accuracy improvements when synthetic videos are added to the training set (e.g., ResNet from 20.46% to 51.15% on Toyota and from 8.66% to 42.98% on NTU), and upward accuracy trends in one-shot and few-shot settings as the number of synthetic videos increases. The paper also introduces the RANDOM People dataset and open-sources code and data.

Significance. If the claimed gains were validated, the approach would be a practical and reusable contribution to low-resource action recognition and to interventions for data diversity. The open-sourced pipeline and the crowd-sourced RANDOM People dataset are concrete assets, and the paper is transparent about several limitations (L1-L4). However, the current experimental validation does not isolate the effect of the pose-transfer pipeline from the effect of simply having more training examples, and the few-shot results lack statistical support. The problem is relevant and the resource is potentially valuable, but the evidence for the central claim is incomplete.

major comments (3)
  1. [Section 5.1 / Table 1] The baseline comparison is confounded by training-set size: the 'original' condition has 225 real videos per class, while the 'original + synthetic' condition has 225 real plus 225 synthetic videos per class, doubling the total number of training examples. The large accuracy gains (ResNet: 20.46 to 51.15 on Toyota and 8.66 to 42.98 on NTU; SlowFast: 38.35 to 55.64 and 26.75 to 36.29) could therefore reflect the general benefit of more data rather than any property of the pose-transfer pipeline. The authors need a matched-quantity control, such as training on 450 real videos per class where the dataset permits it, or on 225 real videos plus 225 realistically augmented real videos, to support the claim that the synthetic data itself is responsible for the improvement.
  2. [Sections 5.2-5.3 / Figures 6-7] The one-shot and few-shot experiments share the same confound: n_real is fixed at 1 or 5 while n_background grows from 0 to 200, so the total training-set size increases along the x-axis. The upward curves are consistent with a generic 'more data helps' effect and do not by themselves demonstrate that the synthetic data generated by the proposed method drives the improvement. Moreover, each plotted point appears to be a single training run on a single randomly selected real sample; no repeated seeds, error bars, or significance tests are reported. To substantiate the few-shot claims, the paper should include multi-seed repeats and matched-quantity controls, such as increasing n_real to 201 or 205 with real videos or with standard augmentations of the real training set.
  3. [Section 3.3 / 4.4] The assumption that pose-transferred videos preserve the action class (e.g., drinking remains drinking) is load-bearing for the method's value, but it is supported only by qualitative examples in Figures 8-11 and the authors' manual review. Without a quantitative measure of action fidelity, it remains possible that the accuracy gains come from appearance or background cues rather than from the action itself, which would limit generalizability to new actions or new pose-transfer backbones. A simple quantitative sanity check, such as running a strong pretrained action classifier on the synthetic clips or measuring per-frame pose alignment against the reference videos, would substantially increase confidence in the method's central mechanism.
minor comments (5)
  1. [Figures 6-7 / Sections 6.2-6.3] The figure captions label one-shot as green and few-shot as red, while the text in Sections 6.2 and 6.3 describes one-shot as red and few-shot as green; the captions and text are inconsistent and should be harmonized.
  2. [References / Section 1 and Section 2] References [20] and [21] are duplicates (same title, same authors, same arXiv identifier) and are cited in close proximity; one should be removed and the citations updated.
  3. [Section 4.1] Please correct the formatting in '8USD' to '$8' and the typo 'derivate artifacts' to 'derivative artifacts'.
  4. [Footnote 2] The footnote uses the word 'ablation' to describe the comparison between training on original-only and original-plus-synthetic data; this is not an ablation because the total training-set size is not held fixed, and the term may mislead readers about what is being compared.
  5. [Section 2.1] In the sentence 'As described Section 3', the preposition 'in' is missing; it should read 'As described in Section 3'.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation: the reported gains are empirical measurements on held-out real test videos; the only self-citation is minor and not load-bearing.

full rationale

The paper contains no mathematical derivation chain to audit: the contribution is an empirical data-generation pipeline (Sections 3-4) and a set of training experiments (Section 5). The central claim is that adding pose-transferred synthetic videos to the training set improves action recognition accuracy on held-out real test videos (Table 1). The test set consists only of original videos ('The testing set, composed only of original videos, included n_test = 50 videos per class'), so the reported gains are not fitted to the evaluation target and do not reduce to the synthetic-data input by construction. The only self-citation, [23] (Knapp and Bohacek), is used in related work to argue that ExAvatar produces more coherent motion than diffusion-based alternatives; that prior human-evaluation study is an independent empirical result and does not presuppose that synthetic training data improves action recognition. No 'uniqueness theorem', ansatz, or renamed known result is imported from the authors' prior work. The main validity concern is a dataset-size confound: the baseline uses 225 real clips per class and the augmented condition uses 225 real plus 225 synthetic clips per class, so the improvement could partly reflect more training examples rather than the pose-transfer pipeline; this is a scientific-control issue, not circularity. Accordingly, no circular step is identified, and the score reflects only the presence of a minor self-citation.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

This is an empirical machine learning paper with no mathematical derivation. The free parameters are experimental settings chosen by hand; they are not fitted to the test set. The axioms are domain assumptions about the pose-transfer backbone, the synthetic-to-real domain gap, and the evaluation protocol. No new physical or conceptual entities are introduced.

free parameters (5)
  • n_real = n_background = 225 videos per class
    Cap placed on per-class training samples, chosen to match the minimum per-class count in Toyota Smarthome and to balance real versus synthetic source ratios. Affects the magnitude of the reported gains.
  • g (backgrounds per video) = 3
    Number of backgrounds sampled for each reference-video and avatar combination. Directly sets the size of the synthetic set (n_T*g*n_A).
  • n_A (avatar identities used) = 15
    Only 15 of the 100 collected identities were used due to compute limits. Results may depend on the specific identities.
  • n_T (reference videos) = 80 (16 classes x 5)
    Hand-selected reference videos per class for pose transfer. Small number may limit motion diversity in synthetic data.
  • training epochs = 5
    All models trained for 5 epochs with learning rate 1e-4; models may be undertrained, which could inflate the relative benefit of synthetic data.
assumptions (4)
  • domain assumption The pose-transfer reenactment preserves the action class of the reference video.
    Central to the method; only validated qualitatively in Appendix 4.
  • domain assumption Synthetic videos composite onto background images without losing action information or introducing a domain gap that prevents transfer to real videos.
    Relied on throughout Sections 3.3 and 5; no quantitative domain-gap analysis.
  • domain assumption The 50-video-per-class test splits are correctly labeled and representative of the action classes.
    Authors created the train and test split; no cross-subject or standard protocol is used.
  • domain assumption The RANDOM People identity set provides sufficient identity and background diversity to benefit underrepresented groups.
    Claimed in Section 8 but not measured in any experiment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Synthetic Human Action Video Data Generation with Pose Transfer." pith.science (2026). https://pith.science/paper/QPE24IAO

@misc{pith2026250609411,
  author       = {Pith},
  title        = {Pith review of: Synthetic Human Action Video Data Generation with Pose Transfer},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QPE24IAO}},
  note         = {Machine review of arXiv:2506.09411}
}
read the original abstract

In video understanding tasks, particularly those involving human motion, synthetic data generation often suffers from uncanny features, diminishing its effectiveness for training. Tasks such as sign language translation, gesture recognition, and human motion understanding in autonomous driving have thus been unable to exploit the full potential of synthetic data. This paper proposes a method for generating synthetic human action video data using pose transfer (specifically, controllable 3D Gaussian avatar models). We evaluate this method on the Toyota Smarthome and NTU RGB+D datasets and show that it improves performance in action recognition tasks. Moreover, we demonstrate that the method can effectively scale few-shot datasets, making up for groups underrepresented in the real training data and adding diverse backgrounds. We open-source the method along with RANDOM People, a dataset with videos and avatars of novel human identities for pose transfer crowd-sourced from the internet.

Figures

Figures reproduced from arXiv: 2506.09411 by the authors.

Figure 1
Figure 1. Examples of synthetic video frames showing [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of our method for synthetic human action video data generation. Avatar example taken from [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Representative frames from the novel human identity videos [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Representative frames from the reference videos [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Representative background images from B in the RANDOM People dataset. Original Synthetic Toyota NTU RGB+D ResNet ✓ 20.46 8.66 ✓ ✓ 51.15 42.98 SlowFast ✓ 38.35 26.75 ✓ ✓ 55.64 36.29 [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Testing accuracy on real videos from the Toyota [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 8
Figure 8. Figure 8: Examples of video frames at t = {0, 1, 2, 3, 4} seconds from the source video (top), taken from Toyota Smarthome, and the target video (bottom), generated by our synthetic data generation method, where the pose alignment is consistent [PITH_FULL_IMAGE:figures/full_fig…
Figure 9
Figure 9. Figure 9: Examples of video frames at t = {0, 1, 2, 3, 4} seconds from the source video (top), taken from Toyota Smarthome, and the target video (bottom), generated by our synthetic data generation method, where the pose alignment is inconsistent [PITH_FULL_IMAGE:figures/full_f…
Figure 10
Figure 10. Figure 10: Examples of video frames at t = {0, 1, 2, 3, 4} seconds from the source video (top), taken from NTU RGB+D dataset, and the target video (bottom), generated by our synthetic data generation method, where the pose alignment is consistent [PITH_FULL_IMAGE:figures/full_f…
Figure 11
Figure 11. Figure 11: Examples of video frames at t = {0, 1, 2, 3, 4} seconds from the source video (top), taken from NTU RGB+D dataset, and the target video (bottom), generated by our synthetic data generation method, where the pose alignment is inconsistent [PITH_FULL_IMAGE:figures/full…
Figure 12
Figure 12. Figure 12: Example video frames illustrating limitation L3 (see Section [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Example video frames illustrating limitation L4 (see Section [PITH_FULL_IMAGE:figures/full_fig_p017_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

66 extracted references · 51 canonical work pages

  1. [1]

    Vivit: A video vision transformer

    Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lu ˇci´c, and Cordelia Schmid. Vivit: A video vision transformer. InProceedings of the IEEE/CVF inter- national conference on computer vision, pages 6836–6846,

  2. [2]

    Is space-time attention all you need for video understanding? InICML, page 4, 2021

    Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? InICML, page 4, 2021. 3

    work page 2021

  3. [3]

    BEDLAM: A synthetic dataset of bodies ex- hibiting detailed lifelike animated motion

    Michael J Black, Priyanka Patel, Joachim Tesch, and Jin- long Yang. BEDLAM: A synthetic dataset of bodies ex- hibiting detailed lifelike animated motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 8726–8737, 2023. 2

    work page 2023

  4. [4]

    Quo vadis, action recognition? a new model and the kinetics dataset

    Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. Inpro- ceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299–6308, 2017. 3

    2017

  5. [5]

    Why can’t i dance in the mall? learning to miti- gate scene bias in action recognition.Advances in Neural Information Processing Systems, 32, 2019

    Jinwoo Choi, Chen Gao, Joseph CE Messou, and Jia-Bin Huang. Why can’t i dance in the mall? learning to miti- gate scene bias in action recognition.Advances in Neural Information Processing Systems, 32, 2019. 8

    work page 2019

  6. [6]

    Openmmlab pose estimation tool- box and benchmark.https://github.com/open- mmlab/mmpose, 2020

    MMPose Contributors. Openmmlab pose estimation tool- box and benchmark.https://github.com/open- mmlab/mmpose, 2020. 4

    work page 2020

  7. [7]

    Toyota smarthome untrimmed: Real-world untrimmed videos for activity detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–1,

    Rui Dai, Srijan Das, Saurav Sharma, Luca Minciullo, Lorenzo Garattoni, Francois Bremond, and Gianpiero Francesca. Toyota smarthome untrimmed: Real-world untrimmed videos for activity detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1–1,

    work page

  8. [8]

    Toyota smarthome: Real-world activities of daily living

    Srijan Das, Rui Dai, Michal Koperski, Luca Minci- ullo, Lorenzo Garattoni, Francois Bremond, and Gianpiero Francesca. Toyota smarthome: Real-world activities of daily living. InProceedings of the IEEE/CVF international con- ference on computer vision, pages 833–842, 2019. 2, 3

    work page 2019

Show all 66 references
  1. [9]

    Fairness and bias mitigation in com- puter vision: A survey.arXiv preprint arXiv:2408.02464,

    Sepehr Dehdashtian, Ruozhen He, Yi Li, Guha Bal- akrishnan, Nuno Vasconcelos, Vicente Ordonez, and Vishnu Naresh Boddeti. Fairness and bias mitigation in com- puter vision: A survey.arXiv preprint arXiv:2408.02464,

    arXiv

  2. [10]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 3

    2010 arXiv

  3. [11]

    A survey on bias in visual datasets.Computer Vision and Image Understanding, 223: 103552, 2022

    Simone Fabbrizzi, Symeon Papadopoulos, Eirini Ntoutsi, and Ioannis Kompatsiaris. A survey on bias in visual datasets.Computer Vision and Image Understanding, 223: 103552, 2022. 8

    2022

  4. [12]

    Pytorchvideo: A deep learning li- brary for video understanding

    Haoqi Fan, Tullie Murrell, Heng Wang, Kalyan Vasudev Al- wala, Yanghao Li, Yilei Li, Bo Xiong, Nikhila Ravi, Meng Li, Haichuan Yang, et al. Pytorchvideo: A deep learning li- brary for video understanding. InProceedings of the 29th ACM international conference on multimedia, pa...

    2021

  5. [13]

    Slowfast networks for video recognition

    Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202–6211, 2019. 3, 6

    2019

  6. [14]

    Black, and Timo Bolkart

    Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3d face model from in-the-wild images, 2021. 4

    2021

  7. [15]

    The” something something” video database for learning and evaluating visual common sense

    Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michal- ski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, et al. The” something something” video database for learning and evaluating visual common sense. InP...

    2017

  8. [16]

    DensePose: Dense human pose estimation in the wild

    Rıza Alp G ¨uler, Natalia Neverova, and Iasonas Kokkinos. DensePose: Dense human pose estimation in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7297–7306, 2018. 3

    2018

  9. [17]

    Can spatiotemporal 3d cnns retrace the history of 2d cnns and im- agenet? InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546–6555, 2018

    Kensho Hara, Hirokatsu Kataoka, and Yutaka Satoh. Can spatiotemporal 3d cnns retrace the history of 2d cnns and im- agenet? InProceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6546–6555, 2018. 3, 6

    2018

  10. [18]

    Animate anyone: Consistent and controllable image- to-video synthesis for character animation

    Li Hu. Animate anyone: Consistent and controllable image- to-video synthesis for character animation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153–8163, 2024. 3

    2024

  11. [19]

    Eldersim: A synthetic data generation platform for human action recognition in eldercare applica- tions.IEEE Access, 11:9279–9294, 2021

    Hochul Hwang, Cheongjae Jang, Geonwoo Park, Junghyun Cho, and Ig-Jae Kim. Eldersim: A synthetic data generation platform for human action recognition in eldercare applica- tions.IEEE Access, 11:9279–9294, 2021. 2

    2021

  12. [21]

    Ai-generated images as data sources: The dawn of synthetic era.arXiv preprint arXiv:2310.01830, 2023

    Michael Jones and Neel Patel. Ai-generated images as data sources: The dawn of synthetic era.arXiv preprint arXiv:2310.01830, 2023. 2

    2023 arXiv

  13. [22]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C. Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross Girshick. Segment anything.arXiv:2304.02643, 2023. 4

    2023 arXiv

  14. [23]

    Can pose transfer models generate realistic human motion?arXiv preprint arXiv:2501.15648, 2025

    Vaclav Knapp and Matyas Bohacek. Can pose transfer models generate realistic human motion?arXiv preprint arXiv:2501.15648, 2025. 3

    2025 arXiv

  15. [24]

    Hmdb: a large video database for human motion recognition

    Hildegard Kuehne, Hueihan Jhuang, Est ´ıbaliz Garrote, Tomaso Poggio, and Thomas Serre. Hmdb: a large video database for human motion recognition. In2011 Inter- national conference on computer vision, pages 2556–2563. IEEE, 2011. 3

    2011

  16. [25]

    Data augmentation techniques for the video question answering task.arXiv preprint arXiv:2008.09849, 2020

    Ziqi Li and Hao Chen. Data augmentation techniques for the video question answering task.arXiv preprint arXiv:2008.09849, 2020. 1

    2008 arXiv

  17. [27]

    Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data.arXiv preprint arXiv:2212.04821, 2022

    Hao Liu and Ting Wu. Promptonomyvit: Multi-task prompt learning improves video transformers using synthetic scene data.arXiv preprint arXiv:2212.04821, 2022. 2

    2022 arXiv

  18. [28]

    Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view syn- thesis

    Wayne Wu Liu, Xian Zhang, Cheng Li, and Chen Change Loy. Liquid warping gan: A unified framework for human motion imitation, appearance transfer and novel view syn- thesis. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 5904–5913, 2019. 3

    2019

  19. [29]

    Smpl: A skinned multi- person linear model

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi- person linear model. InSeminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851–866. 2023. 3

    2023

  20. [30]

    Self-supervised monocu- lar depth estimation on water scenes via specular reflection prior.Digital Signal Processing, 149:104496, 2024

    Zhengyang Lu and Ying Chen. Self-supervised monocu- lar depth estimation on water scenes via specular reflection prior.Digital Signal Processing, 149:104496, 2024. 8

    2024

  21. [31]

    Pose guided person image genera- tion

    Liqian Ma, Xu Jia, Qianru Sun, Bernt Schiele, Tinne Tuyte- laars, and Luc Van Gool. Pose guided person image genera- tion. InAdvances in Neural Information Processing Systems, pages 406–416, 2017. 3

    2017

  22. [32]

    Controllable person image syn- thesis with attribute-decomposed gan

    Yuming Men, Liming Jiang, Jianmin Zhang, Shuaicheng Liu, and Ming-Hsuan Yang. Controllable person image syn- thesis with attribute-decomposed gan. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5084–5093, 2020. 3

    2020

  23. [33]

    Ac- curate 3d hand pose estimation for whole-body 3d human mesh estimation, 2022

    Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Ac- curate 3d hand pose estimation for whole-body 3d human mesh estimation, 2022. 4

    2022

  24. [34]

    Expressive whole-body 3d gaussian avatar, 2024

    Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3d gaussian avatar, 2024. 2, 3, 4

    2024

  25. [35]

    Pytorch: An im- perative style, high-performance deep learning library.Ad- vances in neural information processing systems, 32, 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An im- perative style, high-performance deep learning library.Ad- vances in neural information processing systems, ...

    2019

  26. [36]

    Expressive body capture: 3D hands, face, and body from a single image

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3D hands, face, and body from a single image. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition...

    2019

  27. [37]

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image, 2019. 3, 4

    2019

  28. [38]

    Adaptive stereo depth estimation with multi- spectral images across all lighting conditions.arXiv preprint arXiv:2411.03638, 2024

    Zihan Qin, Jialei Xu, Wenbo Zhao, Junjun Jiang, and Xi- anming Liu. Adaptive stereo depth estimation with multi- spectral images across all lighting conditions.arXiv preprint arXiv:2411.03638, 2024. 8

    2024 arXiv

  29. [39]

    Synthetic-to-real domain adaptation for action recognition: A dataset and baseline performances

    Arun V Reddy, Ketul Shah, William Paul, Rohita Mocharla, Judy Hoffman, Kapil D Katyal, Dinesh Manocha, Celso M De Melo, and Rama Chellappa. Synthetic-to-real domain adaptation for action recognition: A dataset and baseline performances. In2023 IEEE International Conference on ...

    2023

  30. [40]

    Multi-scale attention guided pose transfer

    Prasun Roy, Saumik Bhattacharya, Subhankar Ghosh, and Umapada Pal. Multi-scale attention guided pose transfer. arXiv preprint arXiv:2202.06777, 2022. 3

    2022 arXiv

  31. [41]

    Doubletake: Geometry guided depth estimation

    Mohamed Sayed, Filippo Aleotti, Jamie Watson, Zawar Qureshi, Guillermo Garcia-Hernando, Gabriel Brostow, Sara Vicente, and Michael Firman. Doubletake: Geometry guided depth estimation. InEuropean Conference on Computer Vi- sion, pages 121–138. Springer, 2025. 8

    2025

  32. [42]

    SynthAct: Towards generalizable human action recognition based on synthetic data

    David Schneider, Marco Keller, Zeyun Zhong, Kunyu Peng, Alina Roitberg, J ¨urgen Beyerer, and Rainer Stiefelhagen. SynthAct: Towards generalizable human action recognition based on synthetic data. In2024 IEEE International Con- ference on Robotics and Automation (ICRA), pages ...

    2024

  33. [43]

    NTU RGB+D: A large scale dataset for 3d human activity analysis

    Amir Shahroudy, Jun Liu, Tian-Tsong Ng, and Gang 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. 2, 3, 5, 1

    2016

  34. [44]

    Deformable gans for pose-based human image generation

    Aliaksandr Siarohin, Enver Sangineto, St ´ephane Lathuili`ere, and Nicu Sebe. Deformable gans for pose-based human image generation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3408– 3416, 2018. 3

    2018

  35. [45]

    Ucf101: A dataset of 101 human actions classes from videos in the wild.ArXiv, abs/1212.0402, 2012

    Khurram Soomro, Amir Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild.ArXiv, abs/1212.0402, 2012. 3

    2012 arXiv

  36. [46]

    Synthetic data generation: A hands-on guide in python, 2024

    DataCamp Team. Synthetic data generation: A hands-on guide in python, 2024. 1

    2024

  37. [47]

    Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training.Advances in neural information processing systems, 35:10078–10093, 2022

    Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training.Advances in neural information processing systems, 35:10078–10093, 2022. 3

    2022

  38. [48]

    Synthetic humans for action recognition from un- seen viewpoints.International Journal of Computer Vision, 129(7):2264–2287, 2021

    G ¨ul Varol, Ivan Laptev, Cordelia Schmid, and Andrew Zis- serman. Synthetic humans for action recognition from un- seen viewpoints.International Journal of Computer Vision, 129(7):2264–2287, 2021. 2

    2021

  39. [49]

    Large-scale deep learning optimizations: A comprehensive survey.arXiv preprint arXiv:2111.00856, 2021

    Xin Wang, Kai Xu, et al. Large-scale deep learning optimizations: A comprehensive survey.arXiv preprint arXiv:2111.00856, 2021. 1

    2021 arXiv

  40. [51]

    Generative ai for synthetic data generation: Methods and applications.arXiv preprint arXiv:2403.04190, 2024

    Liyuan Xu and Yifan Zhang. Generative ai for synthetic data generation: Methods and applications.arXiv preprint arXiv:2403.04190, 2024. 2

    2024 arXiv

  41. [52]

    Magicanimate: Temporally consistent human im- age animation using diffusion model

    Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human im- age animation using diffusion model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

    2024

  42. [53]

    Depth any- thing v2, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2, 2024. 4

    2024

  43. [54]

    Learn2augment: Learning to com- posite videos for data augmentation in action recognition

    Wei Zhang and Yao Liu. Learn2augment: Learning to com- posite videos for data augmentation in action recognition. arXiv preprint arXiv:2206.04790, 2022. 1

    2022 arXiv

  44. [55]

    Progressive pose attention transfer for person image generation

    Hao Zhu, Xiaoqian Huang, Hongwei Shi, Xiaoguang Li, Ran He, and Zhenan Wang. Progressive pose attention transfer for person image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2347–2356, 2019. 3

    2019

  45. [56]

    Champ: Controllable and consistent human image animation with 3d parametric guidance.arXiv preprint arXiv:2403.14781, 2024

    Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance.arXiv preprint arXiv:2403.14781, 2024. 3 Synthetic Human Action Video Data Generati...

    2024 arXiv

  46. [57]

    Prolific Participant Instructions As described in Section 4.1, to create theRANDOM Peo- pledataset, we crowd-sourced novel human identity videos using the Prolific data platform. Before entering the record- ing interface and seeing any instructions, the participants were infor...

  47. [58]

    a slow 360 rotation with your hands down

  48. [59]

    a slow 360 rotation with your hands up in a dou- ble L shape as shown below

  49. [60]

    Importantly, the recording must meet the follow- ingcriteria: • Your whole body, head to feet, is visible in the video at all times

    a slow 360 rotation with your hands down. Importantly, the recording must meet the follow- ingcriteria: • Your whole body, head to feet, is visible in the video at all times. • You must be well-lit. • Besides you, no other people, animals, or moving objects appear in the video...

  50. [61]

    In particular, these subsets in- clude: Selected Action Classes: Toyota Smarthome

    Selected Action Classes As described in Section 4, we manually selected a subset of16action classes within the Toyota Smarthome [7] and NTU RGB-D [43] based on the following criteria: (1) Mini- mal Use of External Objects, (2) Consistent Camera Angles, and (3) Distinctive Acti...

  51. [62]

    Maketea.Insertteabag

  52. [63]

    Enter Selected Action Classes: NTU RGB-D

  53. [64]

    kicking something (A24)

  54. [65]

    point to something (A31)

  55. [66]

    put palms together (A39)

  56. [67]

    cross hands in front (A40)

  57. [68]

    Our aim is to provide an intuition for the computational demands of this process and to explain the parameters we chose, which were largely constrained by our computing capacity

    Compute Considerations This appendix section discusses compute considerations surrounding our experimental setup. Our aim is to provide an intuition for the computational demands of this process and to explain the parameters we chose, which were largely constrained by our comp...

  58. [69]

    Qualitative Evaluation t= 0t= 1t= 2t= 3t= 4 Figure 8. Examples of video frames att={0,1,2,3,4}seconds from the source video (top), taken from Toyota Smarthome, and the target video (bottom), generated by our synthetic data generation method, where the pose alignment is consist...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.