Pith. sign in

REVIEW 4 major objections 5 minor 49 references

Diffusion Model-based Activity Completion for AI Motion Capture from Videos

T0 review · 4 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 A small diffusion model can stitch any two human motion clips into one smooth, arbitrarily long sequence.

desk verdict Close HumanMAC variant with two small tweaks and a smaller model, but the central arbitrary-length and cross-action claims are untested and the ablation is confounded. read the letter →

arxiv 2505.21566 v1 pith:J7V4WDDV submitted 2025-05-27 cs.CV cs.LG

Gao Huayu , Huang Tengjiu , Ye Xiaolong , Tsuyoshi Okita This is my paper
classification cs.CVcs.LG
keywords motioncompletiondiffusionmodelhumanpredictioncaptureHuman3.6MgatemoduleIMUdatasynthesisposition-timeembedding
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

The paper proposes MDC-Net, a diffusion-based network that fills the gap between two arbitrary human motion clips, producing a smooth intermediate transition so that the two actions become one continuous, arbitrarily long sequence. The authors argue that such transitions can be predicted from the final frames of the first clip and the initial frames of the second, without seeing the full actions. On the Human3.6M benchmark, MDC-Net reports lower average and final displacement errors than HumanMAC and MDM, while using a smaller model (16.84M parameters vs. 28.40M for HumanMAC). The same generated motions are piped through SMPL mesh estimation and a simulated IMU module to produce synthetic acceleration and angular-velocity signals, which the authors present as a replacement for hardware-based IMU capture.

What carries the argument

The central object is a masked denoising network operating in the DCT domain. Human motion frames are transformed to frequency-domain coefficients, noise is added, and the network denoises them while a mask keeps the observed history and future frames fixed, so the network only fills the middle gap. Two modules carry the argument: a gate module that computes a sigmoid-weighted sum of self-attention and feed-forward outputs, blending global and local features, and a position-time embedding module that encodes temporal location at multiple scales. The implementation uses four stacked modules with skip connections, compared with eight in HumanMAC, which is the source of the smaller parameter count.

What would settle it

Record a subject performing two scripted actions back-to-back with a marker-based motion capture system, feed the same boundary frames used by the paper into MDC-Net, and compare the generated bridge to the recorded bridge with ADE and FDE; if the errors match HumanMAC's on those pairs, or if a joint-angle plausibility check flags the generated frames, the claimed generalization to arbitrary stitching is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that motion completion can be reduced to a masked inpainting problem in the DCT frequency domain: a small diffusion network with a gate module and a position-time embedding can generate the middle 90 frames of a 125-frame window from the first 10 and last 15 observed frames, and this mechanism extends to stitching two arbitrary action clips by padding the gap with the boundary frames. In the quantitative comparison on Human3.6M, MDC-Net reports lower ADE (0.2195 vs 0.2352) and FDE (0.0769 vs 0.0839) than HumanMAC, essentially ties it on MMADE (0.5716 vs 0.5718), and is slightly worse on MMFDE (0.8077 vs 0.7946), while using 16.84M parameters against HumanMAC's 28.40M. The same generated skeleton is then converted to an SMPL mesh and run through a simulated IMU module to output synthetic acceleration and angular velocity for a chosen joint.

Load-bearing premise

The whole product rests on the assumption that a smooth, physically natural bridge between two unrelated actions can be predicted from just the boundary frames of each, since the model is trained only on single-action inpainting and the cross-action case is evaluated without ground truth.

Editorial extensions

If this is right

  • Any two clips from the training distribution can be concatenated into one long clip, so a virtual character can run indefinitely or move from running to sitting without the full sequence being predefined.
  • Because the completion window can be slid and re-applied, the generated sequence can be extended to arbitrary lengths by repeatedly feeding the last frames of the completed motion as the new history.
  • Synthetic IMU signals for any joint, such as the left wrist, can be produced from the generated skeleton through SMPL mesh normals and a simulated IMU module, removing the need for physical inertial sensors.
  • The 16.84M-parameter model reports better ADE, FDE, and MMADE than the 28.40M-parameter HumanMAC on Human3.6M, implying that the gate and time-embedding modules recover the accuracy lost by using half the number of modules.

Reading between the lines

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

  • If the boundary-frame premise holds beyond Human3.6M actions, the same masked-completion recipe could be applied to other periodic or quasi-periodic signals, such as audio or sensor streams, where a smooth transition between two observed segments is needed.
  • The paper's own limitation that some transitions violate physical plausibility suggests that adding explicit biomechanical constraints, like joint-angle limits or foot-ground contact, would be a natural next step for making the arbitrary-length stitching usable in production animation.
  • Because cross-action stitching is not evaluated against ground truth, a direct test would be to record a subject performing two scripted actions back-to-back and compare the generated bridge to the actual recorded bridge; the authors did not report such a comparison.
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 MDC-Net, a diffusion-based motion completion network built on the HumanMAC architecture, augmented with a gate module and a position-time embedding module, to generate transition motion sequences connecting two motion fragments and to extract synthetic IMU data from the resulting motion. Experiments on Human3.6M report lower ADE, FDE, and MMADE than HumanMAC and MDM, a smaller model size (16.84M vs. 28.40M), and qualitative demonstrations of cross-action stitching. The paper claims that MDC-Net accepts motion sequences of any length and generates sequences of arbitrary length, requires less training time than HumanMAC, and supports a pipeline from skeleton to SMPL mesh to simulated IMU signals.

Significance. If the central claims held, MDC-Net would be a lightweight plug-in for extending motion-capture sequences and for generating synthetic IMU training data, which is an appealing practical direction. The paper follows standard Human3.6M train/test splits and reports a concrete parameter-count reduction. However, the significance is currently limited by three factors: the reported metric gains over HumanMAC are marginal and are reported without error bars; the layer-count ablation is confounded with the noise schedule; and the headline capabilities of arbitrary-length generation and cross-action stitching are not supported by any quantitative evaluation. The IMU extraction component is also presented only through qualitative plots, without comparison to real IMU measurements.

major comments (4)
  1. [Section 5.2, Table 1] The claimed quantitative superiority over HumanMAC is not established. The MMADE difference is 0.0002 (0.5716 vs. 0.5718), and MMFDE is worse for MDC-Net (0.8077 vs. 0.7946). No error bars or multiple-run statistics are reported, so the ADE and FDE improvements (0.2195 vs. 0.2352 and 0.0769 vs. 0.0839) may be within run-to-run variation. Additionally, the Conclusion states that MDC-Net outperforms FlowMDM and HumanMAC, but Table 1 does not include FlowMDM. The abstract and conclusion should be tempered or supported with statistically meaningful comparisons.
  2. [Section 5.3, Tables 2 and 3] The layer-count ablation is confounded with the noise schedule. The 4-layer row in Table 2 (ADE 0.2195, FDE 0.0769, MMADE 0.5716, MMFDE 0.8077) is identical to the cosine row of Table 3, and the 8-layer row (ADE 0.5544, FDE 0.3730, MMADE 0.7964, MMFDE 0.8217) is identical to the sqrt row. Therefore the conclusion that 4 layers outperform 8 layers cannot be attributed to the number of layers; the experiment must vary the layer count while keeping the noise schedule fixed.
  3. [Section 5.1 and Contributions] All quantitative experiments use a fixed 125-frame window (10 history frames, 90 prediction frames, and 15 future frames, as stated in Section 5.1). The central contribution claim that MDC-Net 'accepts motion sequences of any length as input and generates motion sequences of arbitrary length' is therefore never tested. No experiment varies the context length or the prediction length, and the cross-action demonstrations in Section 5.4 also use a fixed 125-frame completion. The arbitrary-length claim is unsupported by the reported evidence.
  4. [Section 5.4 and Limitations] The motivating application of stitching two different actions (e.g., Sitting to Walking, Greeting to Phoning) is evaluated only with sampled-frame visualizations and qualitative descriptions. No quantitative metrics are reported for cross-action transitions, and the Limitations section explicitly acknowledges that some generated transitions deviate from realistic human movement and physical laws (e.g., legs extending downward instead of the upper body rising). Since the model is trained to inpaint the middle of a single recorded action, the premise stated in Section 1 that a transition can be predicted from the final frames of H1 and the initial frames of H2 remains a load-bearing but untested assumption. The central product claim collapses if this generalization fails.
minor comments (5)
  1. [Section 6.1, Table 5] The GreetingToPhoning row reports 340 total points, but the sum of used points (35) and generated points (90) is 125; this is internally inconsistent. There is also a typo 'Human3.8M' in the table caption.
  2. [Section 4.2, TimeEmbedding] The position-time embedding module is described only vaguely as 'incorporating information from different time scales'; the exact encoding construction, dimensionality, and how it is combined with the transformer input are not specified, making the module difficult to reproduce.
  3. [Section 5.3, Table 4] The Baseline row in Table 4 has the same values as the HumanMAC row in Table 1 (APD 3.3563, ADE 0.2352, FDE 0.0839, MMADE 0.5718, MMFDE 0.7946), but the text does not state whether the baseline is the unmodified HumanMAC or MDC-Net without the proposed modules; this should be clarified.
  4. [Contributions and Section 5] The contribution bullet claiming 'less training time compared to HumanMAC' is not supported by any reported training-time measurement; please include training-time comparisons or remove the claim.
  5. [Section 6.3] The sentence 'The plot illustrates the acceleration and angular velocity of the left wrist when a person does actions from phoning to walking and from phoning to walking' repeats the same transition pair; presumably one of the two should refer to a different pair of actions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MDC-Net's benchmark results are empirical train/test comparisons, not derivations from its inputs.

full rationale

MDC-Net's central benchmark claims (Section 5.2, Table 1) are empirical train/test comparisons on held-out Human3.6M middle frames, measured with ADE/FDE/MMADE against ground truth. The masking/padding construction in Eq. (3) preserves known H1/H2 frames and lets the diffusion denoiser produce the unknown P; no equation defines the predicted P as a deterministic function of the H1/H2 endpoints, so the reported errors are not forced by construction. The only author self-citation ([1], a pose-estimation paper) appears in a general related-work list and is not load-bearing. The 'arbitrary length' and cross-action stitching claims are only weakly supported—Section 5.1 uses a fixed 125-frame window, Section 5.4 is qualitative, and the Limitations section admits physically implausible transitions—but these are validation/correctness gaps, not circular derivations. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors' prior work, and no known result is merely relabeled.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The reported improvements rest on hand-set hyperparameters (noise schedule, context split, module count), on the domain assumption that local clip boundaries contain enough information to synthesize transitions, and on unvalidated assumptions that the SMPL/MotionBERT/NeuralBody/MATLAB pipeline produces realistic IMU data. The two new architectural components are supported only by in-paper ablations, with no independent external evidence.

free parameters (5)
  • Noise schedule = cosine
    Chosen by comparing four schedules; cosine was selected because it gives the best ADE/FDE/MMADE in Table 3. The choice affects all reported headline results.
  • Context window split = 10 history / 90 prediction / 15 future frames (and 15/20 in Table 5)
    Hand set in Section 5.1; the paper claims splitting the filling equally yields the best results, but no systematic sweep of this split is reported.
  • Number of stacked modules N = 4
    Set to 4 after comparing with 8 in Section 5.3; the 4-layer variant yields the headline metrics, but the comparison is confounded with the noise schedule.
  • Left wrist vertex index = 2208
    Chosen manually in Blender for the IMU sensor placement and described as an experimental choice in Section 6.3.
  • Learning rate and training epochs = 0.0003, 1000 epochs
    Set in Section 5.1 with no sensitivity analysis reported.
assumptions (5)
  • domain assumption The transition between two motion fragments can be predicted from only the final frames of H1 and the initial frames of H2, relying on human motion periodicity.
    Stated in Section 1; this premise makes masked completion the right tool. Cross-action transitions have no ground truth and are only visually evaluated.
  • domain assumption A model trained to inpaint middle segments of single Human3.6M action sequences generalizes to stitching pairs of different action classes.
    The training task masks a contiguous portion of one recorded action, while the deployment task connects two different actions. No quantitative evaluation is provided for the latter in Sections 4.1 and 5.4.
  • domain assumption SMPL, MotionBERT, and NeuralBody provide sufficiently accurate body meshes and vertex normals for simulating IMU signals.
    The pipeline in Sections 6.2 and 6.3 relies on these pretrained models without uncertainty or accuracy checks; the Limitations acknowledge mesh inaccuracies can cause large angular velocity errors.
  • domain assumption The MATLAB IMU module output is a valid proxy for real IMU sensor readings.
    No comparison to real IMU data or noise characteristics is reported in Section 6.3.
  • standard math Standard DDPM forward/reverse process definitions and DCT/iDCT linearity hold.
    Equations (1) through (3) rely on the standard diffusion and DCT machinery from cited works.
invented entities (2)
  • Gate fusion module (sigmoid-weighted sum of attention and FFN outputs)
    purpose: The bias decides which features contribute to the final output, claimed to improve generation accuracy and naturalness.
    Support comes only from the in-paper ablation in Table 4; no independent benchmark or theory predicts when such gating helps.
  • Position-time embedding module
    purpose: Encodes temporal information at multiple time scales to smooth generated motions.
    Ablation results are mixed: adding it alone slightly hurts ADE in Table 4, and no independent evidence is provided for its general benefit.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Diffusion Model-based Activity Completion for AI Motion Capture from Videos." pith.science (2026). https://pith.science/paper/J7V4WDDV

@misc{pith2026250521566,
  author       = {Pith},
  title        = {Pith review of: Diffusion Model-based Activity Completion for AI Motion Capture from Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7V4WDDV}},
  note         = {Machine review of arXiv:2505.21566}
}
read the original abstract

AI-based motion capture is an emerging technology that offers a cost-effective alternative to traditional motion capture systems. However, current AI motion capture methods rely entirely on observed video sequences, similar to conventional motion capture. This means that all human actions must be predefined, and movements outside the observed sequences are not possible. To address this limitation, we aim to apply AI motion capture to virtual humans, where flexible actions beyond the observed sequences are required. We assume that while many action fragments exist in the training data, the transitions between them may be missing. To bridge these gaps, we propose a diffusion-model-based action completion technique that generates complementary human motion sequences, ensuring smooth and continuous movements. By introducing a gate module and a position-time embedding module, our approach achieves competitive results on the Human3.6M dataset. Our experimental results show that (1) MDC-Net outperforms existing methods in ADE, FDE, and MMADE but is slightly less accurate in MMFDE, (2) MDC-Net has a smaller model size (16.84M) compared to HumanMAC (28.40M), and (3) MDC-Net generates more natural and coherent motion sequences. Additionally, we propose a method for extracting sensor data, including acceleration and angular velocity, from human motion sequences.

Figures

Figures reproduced from arXiv: 2505.21566 by the authors.

Figure 1
Figure 1. Human Motion Completion. H1 and H2 are two human motions that can either be different or the same. Using a generative model and inference, we produce an intermediate motion sequence, P, to connect and complete these two motions. FlowMDM [2] and HumanMac [5] are diffusion model-based ap￾proaches designed to handle variations in motion types and transitions. In HumanMac, motion switching is constrained to a fixed leng… view at source ↗
Figure 2
Figure 2. This is the flowchart of MDC-Net. We embed the input data into [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Different padding strategies. We conducted experiments on P using the following four strategies: From first line to fourth line of figure, 1. Filling P with the last frame of H1 and the first frame of H2 respectively; 2. Setting all element of P to zero. 3. Filling all elements of P with the last frame of H1; 4. Filling all elements of P with the first frame of H2. As shown in [PITH_FULL_IMAGE:figures/full_fig_p009… view at source ↗
Figures from the paper (14 more)
Figure 4
Figure 4. Figure 4: Mask.The gray segment represents the sequences after padding, [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Baseline. In the figure, nframes represents the [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: In our module, we introduced a gate structure into a normal [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Since our task focuses on generating smooth and natural mo [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 7
Figure 7. Figure 7: Visualizaiton results of Sqrt strategy. This figure shows the [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: Comparisons using HumanMAC and using MDC-Net. These figures show the visualization results of the transition from sitting to walking using HumanMAC and those using MDC-Net. There are a total of 125 frames. We sampled images from frames 20, 30, 40, 60, 80, 90, 100, 115,…
Figure 9
Figure 9. Figure 9: Visualization results of human motion completion. The actions [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Virtual human structure. It consists 17 joints [PITH_FULL_IMAGE:figures/full_fig_p021_10.png]
Figure 11
Figure 11. Figure 11: This figure describes a person giving directions. The first row of [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: This figure describes the process that the transition from [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 13
Figure 13. Figure 13: This figure describes the process a person is taking a photo of [PITH_FULL_IMAGE:figures/full_fig_p024_13.png]
Figure 14
Figure 14. Figure 14: This figure describes the process that a person is sitting and [PITH_FULL_IMAGE:figures/full_fig_p024_14.png]
Figure 15
Figure 15. Figure 15: This figure describes the process that the transition from [PITH_FULL_IMAGE:figures/full_fig_p025_15.png]
Figure 16
Figure 16. Figure 16: This figure describes the acceleration and angular velocity of a [PITH_FULL_IMAGE:figures/full_fig_p025_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

49 extracted references · 42 canonical work pages

  1. [1]

    Im- provement of Human Action Recognition Using 3D Pose Estima- tion, pages 21–37

    Kohei Adachi, Paula Lago, Tsuyoshi Okita, and Sozo Inoue. Im- provement of Human Action Recognition Using 3D Pose Estima- tion, pages 21–37. Springer Singapore, Singapore, 2021

    work page 2021

  2. [2]

    German Barquero, Sergio Escalera, and Cristina Palmero. Seam- less human motion composition with blended positional encod- Diffusion Model-based Activity Completion for AI Motion Capture from Videos IJABC: International Journal of Activity and Behavior Computing 27 ings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ...

    work page 2024

  3. [3]

    Real- time multi-person 2d pose estimation using part affinity fields

    Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Real- time multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291–7299, 2017

    work page 2017

  4. [4]

    Bodiffusion: Diffusing sparse observations for full-body human motion synthesis

    Angela Castillo, Maria Escobar, Guillaume Jeanneret, Albert Pumarola, Pablo Arbel´ aez, Ali Thabet, and Artsiom Sanakoyeu. Bodiffusion: Diffusing sparse observations for full-body human motion synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 4221–4231, 2023

    work page 2023

  5. [5]

    Humanmac: Masked motion completion for human motion prediction

    Ling-Hao Chen, Jiawei Zhang, Yewen Li, Yiren Pang, Xiaobo Xia, and Tongliang Liu. Humanmac: Masked motion completion for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 9544–9555, 2023

    work page 2023

  6. [6]

    Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose

    Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. InComputer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part VII 16 , pages 769–787. Springer, 2020

    work page 2020

  7. [7]

    Alphapose: Whole- body regional multi-person pose estimation and tracking in real- time

    Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole- body regional multi-person pose estimation and tracking in real- time. IEEE Transactions on Pattern Analysis and Machine In- telligence, 45(6):7157–7173, 2022

    work page 2022

  8. [8]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020

    2020

Show all 49 references
  1. [9]

    Momask: Generative masked modeling of 3d human motions

    Chuan Guo, Yuxuan Mu, Muhammad Gohar Javed, Sen Wang, and Li Cheng. Momask: Generative masked modeling of 3d human motions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1900–1910, 2024

    1900

  2. [10]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the Diffusion Model-based Activity Completion for AI Motion Capture from Videos IJABC: International Journal of Activity and Behavior Computing 28 IEEE confer...

    2016

  3. [11]

    Denoising diffusion probabilistic models

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

    2020

  4. [12]

    Video diffusion models

    Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems , 35:8633– 8646, 2022

    2022

  5. [13]

    Phase-functioned neural networks for character control

    Daniel Holden, Taku Komura, and Jun Saito. Phase-functioned neural networks for character control. ACM Transactions on Graphics (TOG), 36(4):1–13, 2017

    2017

  6. [14]

    Winnet: Wavelet- inspired invertible network for image denoising

    Jun-Jie Huang and Pier Luigi Dragotti. Winnet: Wavelet- inspired invertible network for image denoising. IEEE Trans- actions on Image Processing , 31:4377–4392, 2022

    2022

  7. [15]

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

    2013

  8. [16]

    Alias-free generative adversarial networks

    Tero Karras, Miika Aittala, Samuli Laine, Erik H¨ ark¨ onen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in neural information processing systems, 34:852–863, 2021

    2021

  9. [17]

    Plan, posture and go: Towards open-world text-to-motion generation

    Jinpeng Liu, Wenxun Dai, Chunyu Wang, Yiji Cheng, Yansong Tang, and Xin Tong. Plan, posture and go: Towards open-world text-to-motion generation. arXiv preprint arXiv:2312.14828 , 2023

    2023 arXiv

  10. [18]

    Aggregated multi-gans for controlled 3d human motion prediction

    Zhenguang Liu, Kedi Lyu, Shuang Wu, Haipeng Chen, Yanbin Hao, and Shouling Ji. Aggregated multi-gans for controlled 3d human motion prediction. In Proceedings of the AAAI conference on artificial intelligence , volume 35, pages 2225–2232, 2021

    2021

  11. [19]

    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. In Seminal Graphics Papers: Pushing the Bound- aries, Volume 2 , pages 851–866. 2023. Diffusion Model-based Activity Completion for AI Motion Captur...

    2023

  12. [20]

    3d human mesh estimation from virtual mark- ers

    Xiaoxuan Ma, Jiajun Su, Chunyu Wang, Wentao Zhu, and Yizhou Wang. 3d human mesh estimation from virtual mark- ers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 534–543, 2023

    2023

  13. [21]

    Generative ai for recognizing nurse training activities in skeleton-based video data

    Md Ibrahim Mamun, Shahera Hossain, Md Baharul Islam, and Md Atiqur Rahman Ahad. Generative ai for recognizing nurse training activities in skeleton-based video data. International Journal of Activity and Behavior Computing , 2024(3):1–20, 2024

    2024

  14. [22]

    On hu- man motion prediction using recurrent neural networks

    Julieta Martinez, Michael J Black, and Javier Romero. On hu- man motion prediction using recurrent neural networks. In Pro- ceedings of the IEEE conference on computer vision and pattern recognition, pages 2891–2900, 2017

    2017

  15. [23]

    Nerf: Representing scenes as neural radiance fields for view synthesis

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM , 65(1):99–106, 2021

    2021

  16. [24]

    Posefix: Model-agnostic general human pose refinement network

    Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Posefix: Model-agnostic general human pose refinement network. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7773–7781, 2019

    2019

  17. [25]

    Accurate 3d hand pose estimation for whole-body 3d human mesh estima- tion

    Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Accurate 3d hand pose estimation for whole-body 3d human mesh estima- tion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 2308–2317, 2022

    2022

  18. [26]

    A metaverse: Taxonomy, components, applications, and open challenges

    Sang-Min Park and Young-Gab Kim. A metaverse: Taxonomy, components, applications, and open challenges. IEEE access , 10:4209–4251, 2022

    2022

  19. [27]

    3d human pose estimation in video with temporal convolutions and semi-supervised training

    Dario Pavllo, Christoph Feichtenhofer, David Grangier, and Michael Auli. 3d human pose estimation in video with temporal convolutions and semi-supervised training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 7753–7762, 2019

    2019

  20. [28]

    Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans

    Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and...

    2021

  21. [29]

    Recognition of endotracheal suctioning activities: A feature extraction and ensemble learning approach based on pose esti- mation data

    Hoang Khang Phan, Tu Nhat Khang Nguyen, Truong Vi Bui, Khuong Cong Duy Nguyen, Tuan Phong Nguyen, and Nhat Tan Le. Recognition of endotracheal suctioning activities: A feature extraction and ensemble learning approach based on pose esti- mation data. In 2024 International Conf...

    2024

  22. [30]

    Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2

    Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 3626–3636, 2022

    2022

  23. [31]

    Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  24. [32]

    Deepphase: Periodic autoencoders for learning motion phase manifolds

    Sebastian Starke, Ian Mason, and Taku Komura. Deepphase: Periodic autoencoders for learning motion phase manifolds. ACM Transactions on Graphics (TOG) , 41(4):1–13, 2022

    2022

  25. [33]

    Deep high- resolution representation learning for human pose estimation

    Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high- resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5693–5703, 2019

    2019

  26. [34]

    Augmented reality and robotics: A sur- vey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces

    Ryo Suzuki, Adnan Karim, Tian Xia, Hooman Hedayati, and Nicolai Marquardt. Augmented reality and robotics: A sur- vey and taxonomy for ar-enhanced human-robot interaction and robotic interfaces. In Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems , ...

    2022

  27. [35]

    From screen to re- ality: Exploring the evolution and integration of motion capture technology for virtual digital humans

    JiChu Tang, KiHong Kim, and KaiXing Wang. From screen to re- ality: Exploring the evolution and integration of motion capture technology for virtual digital humans. 2024

    2024

  28. [36]

    Human motion diffusion model

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

    2022 arXiv

  29. [37]

    Learning from synthetic humans

    Gul Varol, Javier Romero, Xavier Martin, Naureen Mahmood, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning from synthetic humans. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 109–117, 2017. Diffusion Model-based Activity Co...

    2017

  30. [38]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information pro- cessing systems, 30, 2017

    2017

  31. [39]

    Diffusion- phase: Motion diffusion in frequency domain

    Weilin Wan, Yiming Huang, Shutong Wu, Taku Komura, Wen- ping Wang, Dinesh Jayaraman, and Lingjie Liu. Diffusion- phase: Motion diffusion in frequency domain. arXiv preprint arXiv:2312.04036, 2023

    2023 arXiv

  32. [40]

    Free- man: Towards benchmarking 3d human pose estimation in the wild

    Jiong Wang, Fengyu Yang, Wenbo Gou, Bingliang Li, Danqi Yan, Ailing Zeng, Yijun Gao, Junle Wang, and Ruimao Zhang. Free- man: Towards benchmarking 3d human pose estimation in the wild. arXiv preprint arXiv:2309.05073 , 2023

    2023 arXiv

  33. [41]

    Actformer: A gan-based transformer to- wards general action-conditioned 3d human motion generation

    Liang Xu, Ziyang Song, Dongliang Wang, Jing Su, Zhicheng Fang, Chenjing Ding, Weihao Gan, Yichao Yan, Xin Jin, Xi- aokang Yang, et al. Actformer: A gan-based transformer to- wards general action-conditioned 3d human motion generation. In Proceedings of the IEEE/CVF Internation...

    2023

  34. [42]

    Using ai motion capture systems to capture race walking technol- ogy at a race scene: A comparative experiment

    Dongtao Zhang, Zhongqiu Ji, Guiping Jiang, and Weiwei Jiao. Using ai motion capture systems to capture race walking technol- ogy at a race scene: A comparative experiment. Applied Sciences, 13(1):113, 2022

    2022

  35. [43]

    Mode- adaptive neural networks for quadruped motion control

    He Zhang, Sebastian Starke, Taku Komura, and Jun Saito. Mode- adaptive neural networks for quadruped motion control. ACM Transactions on Graphics (TOG) , 37(4):1–11, 2018

    2018

  36. [44]

    Motiondiffuse: Text- driven human motion generation with diffusion model

    Mingyuan Zhang, Zhongang Cai, Liang Pan, Fangzhou Hong, Xinying Guo, Lei Yang, and Ziwei Liu. Motiondiffuse: Text- driven human motion generation with diffusion model. arXiv preprint arXiv:2208.15001, 2022

    2022 arXiv

  37. [45]

    Dstformer: 3d human pose estimation with a dual-scale spatial and temporal transformer network

    Shaokun Zhang, Xinde Li, Chuanfei Hu, Jianping Xu, and Huap- ing Liu. Dstformer: 3d human pose estimation with a dual-scale spatial and temporal transformer network. In 2024 International Conference on Advanced Robotics and Mechatronics (ICARM) , pages 484–489. IEEE, 2024

    2024

  38. [46]

    Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Zilong Dong, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d Diffusion Model-based Activity Completion for AI Motion Capture from Videos IJABC: International Journ...

    2024

  39. [47]

    Motionbert: A unified perspective on learning human motion representations

    Wentao Zhu, Xiaoxuan Ma, Zhaoyang Liu, Libin Liu, Wayne Wu, and Yizhou Wang. Motionbert: A unified perspective on learning human motion representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15085–15099, 2023

    2023

  40. [48]

    Human motion generation: A survey

    Wentao Zhu, Xiaoxuan Ma, Dongwoo Ro, Hai Ci, Jinlu Zhang, Jiaxin Shi, Feng Gao, Qi Tian, and Yizhou Wang. Human motion generation: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence , 2023

    2023

  41. [49]

    Sensor data augmentation from skeleton pose sequences for improving human activity recogni- tion

    Parham Zolfaghari, Vitor Fortes Rey, Lala Ray, Hyun Kim, Sungho Suh, and Paul Lukowicz. Sensor data augmentation from skeleton pose sequences for improving human activity recogni- tion. In 2024 International Conference on Activity and Behavior Computing (ABC) , pages 1–8, 2024...

    2024

Pith tools

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