Pith. sign in

REVIEW 4 major objections 5 minor 39 references

Motion Diffusion Autoencoders: Enabling Attribute Manipulation in Human Motion Demonstrated on Karate Techniques

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read This paper claims to be the first to manipulate individual attributes of human motion data, changing a karate technique or an athlete's grade while preserving everything else.

desk verdict A solid pose-representation and diffusion-autoencoder adaptation for karate motion, but the central 'first success' manipulation claim is undercut by a circular lambda-selection and a FID table that shows the opposite. read the letter →

arxiv 2501.18729 v2 pith:PA2S3UN4 submitted 2025-01-30 cs.CV cs.LG

Anthony Richardson , Felix Putze This is my paper · ORCID
classification cs.CVcs.LG
keywords DiffusionAutoencoderAttributeManipulationLinearSeparabilityHumanMotionPoseRepresentationEmbeddingSpaceKarate
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

Attribute manipulation means changing one property of a data point while leaving everything else alone, and this paper claims to be the first to do it successfully for human motion: in motion-capture karate recordings, it can change the executed technique or the athlete's grade while preserving the other attribute and the stochastic details of the performance. The authors build a motion diffusion autoencoder that separates each motion into a semantic embedding, produced by a transformer encoder, and a stochastic embedding, produced by a diffusion process; manipulation consists of moving the semantic embedding along a learned linear direction for the target attribute. If the claim holds, it matters because it makes human motion editable without re-recording, with applications in sports training, kinesiology, and animation, and because it extends the diffusion-autoencoder paradigm from images to time series with time-variant attributes. The demonstration on the Kyokushin karate dataset shows grade manipulation working in both directions and technique manipulation working for three of the five techniques, with reverse punches and spinning back kicks resisting the change.

What carries the argument

The load-bearing mechanism is the two-embedding autoencoder, called MoDiffAE: a transformer encoder $f_\psi(x_0)=z$ creates the semantic embedding, a deterministic diffusion process creates the stochastic embedding $x_T$, and a Denoising Diffusion Implicit Model (DDIM) decoder conditioned on both embeddings reconstructs the motion. The manipulation itself uses a linear classifier trained on frozen semantic embeddings; its per-attribute weight vectors serve as linear directions of change, and a guidance mechanism interpolates the manipulation strength $\lambda$ between $0$ and the convergence point $\lambda_{\max}$, selecting the embedding whose predicted attributes best match the targets. The supporting pose representation maps each adjacent-marker pair to an axis-angle rotation, embeds it in the Stiefel manifold (a continuous rotation parameterization) by dropping the last column of the rotation matrix, and reconstructs the skeleton iteratively with a Gram-Schmidt-like inverse that treats bone lengths as constants.

What would settle it

Using the published code and fixed data splits, count how often a grade manipulation changes the technique predicted by the paper's own linear classifier; if that rate is comparable to the classifier's baseline error on unmanipulated motions, the claimed preservation of untargeted attributes fails.

Watch

Extended reading notes

Core claim

The central discovery is that a semantic embedding space learned by a transformer encoder, when paired with a diffusion-based stochastic embedding, is approximately linearly separable with respect to high-level motion attributes, so the weight vectors of a linear classifier become the directions in which attributes change. Moving a motion's semantic embedding along one of these directions while keeping the stochastic embedding fixed produces a reconstructed motion that exhibits the target attribute and preserves the source attribute and the execution-specific variation. To make this possible on marker-based motion capture, the paper introduces a continuous, rotation-based pose representation built on the Stiefel manifold and a chain of adjacent markers, with a reversible mapping that preserves the individual's anatomy and a reconstruction error of 1.05 cm. The paper presents this as the first success at attribute manipulation in human motion, with the explicit boundary that technique changes involving reverse punches and spinning back kicks frequently fail.

Load-bearing premise

The entire editing scheme depends on the semantic embedding capturing exactly what should change and the stochastic embedding preserving exactly what should stay the same, a clean separation the paper concedes is not guaranteed.

Editorial extensions

If this is right

  • Because the stochastic embedding is never changed, manipulated motions keep the idiosyncratic variations of the original recording rather than collapsing into template-like outputs.
  • Any attribute that a linear classifier can read from the semantic embedding becomes editable with the same procedure, without retraining the generative model.
  • The released fixed data splits, preprocessing, and linear-separability measurements constitute a first benchmark for human motion manipulation, giving future methods a common comparison.
  • The current boundary of the claim is explicit: technique manipulations involving reverse punches or spinning back kicks often fail to alter the technique, so the linear-direction method does not cover all attributes equally.

Reading between the lines

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

  • The marker-chain-specific pose representation leaves open whether a chain-independent version could preserve anatomy across different marker sets; if such a version exists, the method would extend beyond the current dataset.
  • The reported failure on reverse punches and spinning back kicks suggests those techniques may occupy a non-linear region of the semantic embedding; analysing the neighbourhood structure of those classes would test whether the linear-direction assumption is the bottleneck.
  • Applying the same two-embedding scheme to attributes that vary within a motion, such as changing only the first segment of a technique, would stress the paper's claim that the semantic encoder captures time-variant semantics using temporal context.
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

4 major / 5 minor

Summary. The paper proposes Motion Diffusion Autoencoders (MoDiffAE), an adaptation of diffusion autoencoders to human motion, and claims the first successful attribute manipulation on human motion data. The model encodes a motion into a semantic embedding via a transformer encoder and a stochastic embedding via a diffusion process; manipulation is performed by moving the semantic embedding along a direction obtained from a trained linear classifier. The paper also introduces a continuous, rotation-based pose representation that preserves anatomical details, with an average reconstruction error of 1.05 cm. Evaluation consists of an embedding-space linear separability analysis, an FID-based comparison (Table 1), and a qualitative inspection of manipulated motions. The authors report successful technique manipulation for three of five karate techniques and grade manipulation in both directions.

Significance. If the central claim holds, this would be a meaningful advance for human motion modeling, enabling controllable edits of motion attributes while preserving identity-specific stochastic variations. The paper contributes a novel pose representation with a differentiable reconstruction that preserves skeleton anatomy, and it makes code and data publicly available, which is commendable for reproducibility. The embedding-space linearity results (average recall 0.789 for techniques, MAE 0.146 for grades) are suggestive, and the qualitative observations for front kicks and roundhouse kicks, as well as grade changes, are plausible. However, the quantitative support for the manipulation claim is currently weak: the only distribution-based metric (Table 1) shows a systematic increase in FID to all reference groups after manipulation, and the per-sample guidance strength is selected using the same classifier that is later used to validate the manipulation, creating a circular evaluation.

major comments (4)
  1. [§8, Table 1] The per-sample guidance strength lambda is chosen by minimizing the distance between the linear classifier's predictions and the target attributes, and the same linear classifier is then used to support the embedding-space linearity claims. This makes the reported post-manipulation classifier-based success partly circular: the selected lambda is optimized to satisfy that classifier. Independent evidence, such as a held-out classifier or a metric that does not involve the direction-finding classifier, is needed to confirm that the manipulation genuinely changes the attribute.
  2. [Table 1] The FID results in Table 1 directly contradict the claim that manipulation moves samples closer to the target distribution. For the HRK-to-FK manipulation, the FID to the target FK group increases from 33.16 to 85.84, and the FID to every other technique and grade group also increases substantially (e.g., HRK from 26.69 to 86.71). This is a systematic off-manifold shift rather than a targeted change, and the authors' acknowledgment that FID is unreliable with fewer than 50 samples does not explain why manipulated samples become more distant from all reference groups. The authors should provide an alternative quantitative metric that supports the on-manifold nature of the manipulated motions.
  3. [§9.2 and §10] The qualitative evaluation is selective and, by the authors' own summary in the conclusion, only demonstrates successful technique manipulation for three out of five techniques. The examples in Figure 7 are described as 'frequent observations,' but no counts or systematic scoring criteria are reported, and the failure cases for reverse punches and spinning back kicks are not analyzed. If the claimed 'first success' is to be convincing, the evaluation needs a protocol with predefined success criteria, a full reporting of success/failure rates across all test samples, and a discussion of failure modes.
  4. [§7.2, §8] The manipulation procedure assumes a clean separation between semantic content and stochastic variation: the semantic embedding is moved while the stochastic embedding is frozen. However, Section 7.2 states that 'there is no guarantee that this will result in purely semantic and stochastic embeddings.' The paper provides no quantitative test of whether the stochastic embedding actually preserves the untargeted variations after manipulation. For example, one could compare the reconstruction of the original motion using the original stochastic embedding with the reconstruction after manipulation, measuring the similarity of low-level details (e.g., per-joint velocity profiles or local pose differences). Without such a check, the preservation claim that defines attribute manipulation remains unsupported.
minor comments (5)
  1. [§1] There is a typo in the introduction: 'can can degrade' should read 'can degrade.'
  2. [§7.2] The word 'stoachstic' appears in the sentence 'there is no guarantee that this will result in purely semantic and stoachstic embeddings'; it should be 'stochastic.'
  3. [§10] The phrase 'marker sets' is missing an apostrophe: 'use-cases with different markers sets' should be 'use-cases with different marker sets.'
  4. [§6.3] The reconstruction error of 1.05 cm is reported as an average, but the distribution of errors (e.g., standard deviation, per-joint maxima) is not given. Reporting the maximum accumulated error would help assess the worst-case distortion during manipulation.
  5. [§9.1] The UMAP projection in Figure 6 is described as showing 'separation of low and high grades inside of each technique cluster,' but the figure is small and the grade coloring is not visible in the provided text. The authors should ensure that the figure is legible and that the claim is supported by the visual evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity found; the lambda-selection procedure is a control mechanism, not an evaluation, and the central claim rests on external FID and expert-based evaluation.

full rationale

The paper's core derivation chain is not circular. The semantic and stochastic encoders are trained with a reconstruction loss on motion data, and the linear classifier is trained on held-out validation embeddings to measure linear separability; this is an independent, falsifiable check. The per-sample lambda in Section 8 is selected by minimizing the linear classifier's distance to the target attributes, which is a guidance mechanism rather than a predictive claim. The paper does not use the classifier's post-manipulation outputs as evidence of success; instead, the success claim is supported by qualitative expert evaluation (Section 9.2) and an FID-based comparison (Section 9.1), both of which are external to the classifier. Notably, Table 1 shows FID worsening after manipulation, which contradicts any claim that the manipulation is forced by construction. The pose representation is reversible by design, but that is a self-consistency property, not a circular prediction. The paper contains no load-bearing self-citations; references to Preechakul et al. are to independent prior work. The acknowledged limitations, such as the lack of a guarantee of purely semantic and stochastic embeddings, are stated as caveats rather than used to derive results. Therefore, no circular step reduces the paper's claims to their inputs.

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

The central contribution rests on several assumptions borrowed from prior work or chosen by hand, including constant bone lengths, semantic/stochastic disentanglement, and linear classifier directions. There are no invented entities.

free parameters (4)
  • Loss weights phi_pos, phi_foot, phi_vel = not stated in main text
    Equation 23 weights the geometric losses; values are chosen by hand and not derived.
  • Guidance strength lambda = per-sample value between 0 and lambda_max
    Section 8 selects lambda by interpolating and scoring with the linear classifier; it is a free tuning parameter, not a prediction.
  • Z-score outlier thresholds = not specified in main text
    Section 5 preprocessing uses statistical criteria to remove or correct recordings; thresholds are author-chosen.
  • Diffusion schedule and step count T = not stated in main text
    The noise schedule and number of diffusion steps are hyperparameters detailed in Appendix 4.
assumptions (4)
  • domain assumption Bone lengths are constant during a motion
    Section 6.3: the pose representation stores average adjacent-marker distances and treats them as constants; supported by a measured mean standard deviation of 0.35 cm on this dataset only.
  • domain assumption Semantic and stochastic embeddings are disentangled
    Section 7.2: the architecture relies on the semantic encoder capturing high-level semantics and the diffusion path capturing low-level variations; the paper explicitly says there is no guarantee of pure separation.
  • standard math DDIM reverse diffusion with the given noise schedule is valid
    Section 7.1: the decoder uses the standard DDIM formulation from Song et al. and Ho et al.
  • domain assumption Linear classifier weight vectors are valid manipulation directions
    Section 8: manipulation moves embeddings along classifier weights; the paper validates approximate linear separability but the direction is only as good as the classifier.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Motion Diffusion Autoencoders: Enabling Attribute Manipulation in Human Motion Demonstrated on Karate Techniques." pith.science (2026). https://pith.science/paper/PA2S3UN4

@misc{pith2026250118729,
  author       = {Pith},
  title        = {Pith review of: Motion Diffusion Autoencoders: Enabling Attribute Manipulation in Human Motion Demonstrated on Karate Techniques},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PA2S3UN4}},
  note         = {Machine review of arXiv:2501.18729}
}
read the original abstract

Attribute manipulation deals with the problem of changing individual attributes of a data point or a time series, while leaving all other aspects unaffected. This work focuses on the domain of human motion, more precisely karate movement patterns. To the best of our knowledge, it presents the first success at manipulating attributes of human motion data. One of the key requirements for achieving attribute manipulation on human motion is a suitable pose representation. Therefore, we design a novel continuous, rotation-based pose representation that enables the disentanglement of the human skeleton and the motion trajectory, while still allowing an accurate reconstruction of the original anatomy. The core idea of the manipulation approach is to use a transformer encoder for discovering high-level semantics, and a diffusion probabilistic model for modeling the remaining stochastic variations. We show that the embedding space obtained from the transformer encoder is semantically meaningful and linear. This enables the manipulation of high-level attributes, by discovering their linear direction of change in the semantic embedding space and moving the embedding along said direction. All code and data is made publicly available.

Figures

Figures reproduced from arXiv: 2501.18729 by the authors.

Figure 1
Figure 1. Manipulating the technique and skill level of karate athletes with our proposed Motion Diffusion Autoencoder. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Number of samples for each technique per grade. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Extracting the angle 𝜃 for two markers 𝑎 and 𝑏. it possible to formulate a reversible mapping from coordinates into a continuous pose representation that preserves anatomical details. 6.1 From Coordinates to Features Given the original coordinate representation, an axis angle 𝜔 for two adjacent markers 𝑎 and 𝑏 is calculated by 𝜔 = 𝑎 ′ × 𝑏 ′ ||𝑎 ′ × 𝑏 ′ ||2 | {z } axis 𝑘 · atan2  𝑎 ′ × 𝑏 ′ · 𝑎 ′ × 𝑏 ′ ||𝑎 ′ × 𝑏 ′ ||… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Overview of the MoDiffAE architecture. enables attribute manipulation on sequential data, while maintain￾ing the high reconstruction capabilities of diffusion models. In other words, it seeks to extract a semantically meaningful and decodable representation. Preechakul…
Figure 5
Figure 5. Figure 5: Grade (left) and technique (right) confusion matrices for validation data. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: UMAP projection of semantic embeddings of the train data with visualized grades (left) and techniques (right). [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Exemplary technique and grade manipulations. Timestamps are provided below each frame. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 27 canonical work pages

  1. [1]

    Zahid Akhtar, Murshida Rahman Mouree, and Dipankar Dasgupta. 2020. Utility of deep learning features for facial attributes manipulation detection. In 2020 IEEE International Conference on Humanized Computing and Communication with Artificial Intelligence (HCCAI). IEEE, 55–60

    work page 2020

  2. [2]

    Sadegh Aliakbarian, Fatemeh Sadat Saleh, Mathieu Salzmann, Lars Petersson, and Stephen Gould. 2020. A stochastic conditioning scheme for diverse human motion prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5223–5232

    work page 2020

  3. [3]

    Roger Bartlett. 2008. Movement variability and its implications for sports sci- entists and practitioners: an overview. International journal of sports science & coaching 3, 1 (2008), 113–124

    work page 2008

  4. [4]

    Helmi Chaabene, Younés Hachana, Emerson Franchini, Bessem Mkaouer, and Karim Chamari. 2012. Physical and physiological profile of elite karate athletes. Sports medicine 42 (2012), 829–843

    work page 2012

  5. [5]

    Helmi Chaabene, Yassine Negra, Laura Capranica, Olaf Prieske, and Urs Granacher. 2019. A needs analysis of karate kumite with recommendations for performance testing and training. Strength & Conditioning Journal 41, 3 (2019), 35–46

    work page 2019

  6. [6]

    Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. 2023. Stablevideo: Text- driven consistency-aware diffusion video editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision . 23040–23050

    2023

  7. [7]

    Nikolas De Giorgis, Enrico Puppo, Paolo Alborno, and Antonio Camurri. 2019. Evaluating Movement Quality Through Intrapersonal Synchronization. IEEE Transactions on Human-Machine Systems 49, 4 (2019), 304–313. https://doi.org/ 10.1109/THMS.2019.2912498

    work page arXiv 2019

  8. [8]

    Prafulla Dhariwal and Alexander Nichol. 2021. Diffusion models beat gans on image synthesis. Advances in neural information processing systems 34 (2021), 8780–8794

    2021

Show all 39 references
  1. [9]

    David E Dobbs. 2023. How to Motivate and Remember the Law of Cosines, the Law of Sines and the Law of Tangents and the Connections Between these Laws. Moroccan Journal of Algebra and Geometry with Applications (2023), 1–35

    2023

  2. [10]

    Adham Elarabawy, Harish Kamath, and Samuel Denton. 2022. Direct inversion: Optimization-free text-driven real image editing with diffusion models. arXiv preprint arXiv:2211.07825 (2022)

    2022 arXiv

  3. [11]

    Ruoyu Feng, Wenming Weng, Yanhui Wang, Yuhui Yuan, Jianmin Bao, Chong Luo, Zhibo Chen, and Baining Guo. 2024. Ccedit: Creative and controllable video editing via diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 6712–6722

    2024

  4. [12]

    F Sebastian Grassia. 1998. Practical parameterization of rotations using the exponential map. Journal of graphics tools 3, 3 (1998), 29–48

    1998

  5. [13]

    Chuan Guo, Xinxin Zuo, Sen Wang, Shihao Zou, Qingyao Sun, Annan Deng, Minglun Gong, and Li Cheng. 2020. Action2motion: Conditioned generation of 3d human motions. In Proceedings of the 28th ACM International Conference on Multimedia. 2021–2029

    2020

  6. [14]

    Fei Han, Brian Reily, William Hoff, and Hao Zhang. 2017. Space-time representa- tion of people based on 3D skeletal data: A review. Computer Vision and Image Understanding 158 (2017), 85–105

    2017

  7. [15]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. GANs trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30 (2017)

    2017

  8. [16]

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33 (2020), 6840–6851

    2020

  9. [17]

    Mohamed E Hussein, Marwan Torki, Mohammad A Gowayyed, and Motaz El- Saban. 2013. Human action recognition using a temporal hierarchy of covariance descriptors on 3d joint locations. In Twenty-third international joint conference on artificial intelligence

    2013

  10. [18]

    Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. 2013. Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE transactions on pattern analysis and machine intelligence 36, 7 (2013), 1325–1339

    2013

  11. [19]

    Yanli Ji, Feixiang Xu, Yang Yang, Fumin Shen, Heng Tao Shen, and Wei-Shi Zheng

  12. [20]

    Tero Karras, Samuli Laine, and Timo Aila. 2019. A Style-Based Generator Ar- chitecture for Generative Adversarial Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    2019

  13. [21]

    Gyeongman Kim, Hajin Shim, Hyunsu Kim, Yunjey Choi, Junho Kim, and Eunho Yang. 2023. Diffusion Video Autoencoders: Toward Temporally Consistent Face Video Editing via Disentangled Video Encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognit...

    2023

  14. [22]

    Jing Lin, Ailing Zeng, Shunlin Lu, Yuanhao Cai, Ruimao Zhang, Haoqian Wang, and Lei Zhang. 2024. Motion-x: A large-scale 3d expressive whole-body human motion dataset. Advances in Neural Information Processing Systems 36 (2024)

    2024

  15. [23]

    Zhenguang Liu, Shuang Wu, Shuyuan Jin, Shouling Ji, Qi Liu, Shijian Lu, and Li Cheng. 2022. Investigating pose representations and motion contexts modeling for 3D motion prediction. IEEE transactions on pattern analysis and machine intelligence 45, 1 (2022), 681–697

    2022

  16. [24]

    Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Gerard Pons-Moll, and Michael J Black. 2019. AMASS: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF international conference on computer vision . 5442– 5451

    2019

  17. [25]

    Christian Mandery, Ömer Terlemez, Martin Do, Nikolaus Vahrenkamp, and Tamim Asfour. 2015. The KIT whole-body human motion database. In 2015 International Conference on Advanced Robotics (ICAR) . IEEE, 329–336

    2015

  18. [26]

    Georgios Pavlakos, Luyang Zhu, Xiaowei Zhou, and Kostas Daniilidis. 2018. Learning to estimate 3D human pose and shape from a single color image. In Proceedings of the IEEE conference on computer vision and pattern recognition . 459–468

    2018

  19. [27]

    Matthias Plappert, Christian Mandery, and Tamim Asfour. 2016. The KIT motion- language dataset. Big data 4, 4 (2016), 236–252

    2016

  20. [28]

    Konpat Preechakul, Nattanat Chatthee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. 2022. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10619–10629

    2022

  21. [29]

    Manuel M Probst, Richard Fletcher, and Dayna S Seelig. 2007. A Comparison of Lower-Body Flexibility, Strength, And Knee Stability between KArate Athletes and Active Controls. The Journal of Strength & Conditioning Research 21, 2 (2007), 451–455

    2007

  22. [30]

    Ashutosh Saxena, Justin Driemeyer, and Andrew Y Ng. 2009. Learning 3-d object orientation from images. In 2009 IEEE International conference on robotics and automation. IEEE, 794–800

    2009

  23. [31]

    Wei Shen and Rujie Liu. 2017. Learning residual images for face attribute ma- nipulation. In Proceedings of the IEEE conference on computer vision and pattern recognition. 4030–4038

    2017

  24. [32]

    Mingyi Shi, Kfir Aberman, Andreas Aristidou, Taku Komura, Dani Lischinski, Daniel Cohen-Or, and Baoquan Chen. 2020. Motionet: 3d human motion recon- struction from monocular video with skeleton consistency. ACM Transactions on Graphics (TOG) 40, 1 (2020), 1–15

    2020

  25. [33]

    Jiaming Song, Chenlin Meng, and Stefano Ermon. 2020. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502 (2020)

    2020 arXiv

  26. [34]

    Agnieszka Szczkesna, Monika Błaszczyszyn, and Magdalena Pawlyta. 2021. Op- tical motion capture dataset of selected techniques in beginner and advanced Kyokushin karate athletes. Scientific Data 8, 1 (2021), 13

    2021

  27. [35]

    Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Daniel Cohen-Or, and Amit H Bermano. 2022. Human motion diffusion model. arXiv preprint arXiv:2209.14916 (2022)

    2022 arXiv

  28. [36]

    Vicon. 2023. Full body modeling with Plug-in Gait. https://docs.vicon.com/ display/Nexus212/Full+body+modeling+with+Plug-in+Gait. Accessed: 2023-11- 20

    2023

  29. [37]

    Yilin Wang, Suhang Wang, Guojun Qi, Jiliang Tang, and Baoxin Li. 2018. Weakly supervised facial attribute manipulation via deep adversarial network. In 2018 IEEE Winter Conference on Applications of Computer Vision (W ACV). IEEE, 112– 121

    2018

  30. [38]

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. 2019. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition . 5745–5753

    2019

  31. [2019]

    arXiv preprint arXiv:1904.10681 (2019)

    A large-scale varying-view rgb-d action dataset for arbitrary-view human action recognition. arXiv preprint arXiv:1904.10681 (2019)

    2019 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.