Pith. sign in

REVIEW 4 major objections 6 minor 36 references

Spatio-Temporal Multi-Subgraph GCN for 3D Human Motion Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

Pith's one-line read This paper claims that a dual-branch graph convolutional network with cross-domain consistency losses and multi-subgraph averaging reduces mean per-joint position error below seven baselines on two standard 3D human motion prediction…

desk verdict Reasonable incremental GCN architecture, but the headline gains are weakened by hyperparameter selection on the test set and missing error bars. read the letter →

arxiv 2501.00317 v1 pith:D7T25USO submitted 2024-12-31 cs.CV cs.LG

classification cs.CVcs.LG
keywords humanmotionpredictiongraphconvolutionalnetworkspatio-temporalmodelingmulti-subgraphlearningconsistencyconstraintMPJPEHuman3.6MCMUMocap
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 proposes STMS-GCN, a graph-convolutional network for 3D human motion prediction that models temporal and spatial dependencies in separate branches and then lets the branches teach each other. The authors claim that this decoupling, plus a multi-subgraph scheme that averages several learned graphs, produces lower mean per-joint position error (MPJPE) than seven published baselines on the Human3.6M and CMU Mocap benchmarks. On H3.6M, the reported average MPJPE is 33.80, with reductions of 3.71% at 80 ms and 3.13% at 160 ms; on CMU Mocap the average is 32.43. If correct, the claim matters because it suggests that explicit cross-domain consistency between temporal and spatial features is an effective way to improve motion forecasting without larger models or extra data.

What carries the argument

The central object is the Spatio-Temporal Multi-Subgraph Block (STMSB), a repeated block with $L$ layers that contains two orthogonal GCN branches and $K$ subgraph convolutions per branch. The block's work is to produce complementary temporal and spatial representations and to enforce agreement between them; the load-bearing identity is the combined loss $L = L_1 + \lambda(L_{ST} + L_{con}^S + L_{con}^T)$, where $L_1$ is the MPJPE prediction loss, $L_{ST}$ aligns temporal and spatial branch outputs at each scale, and the $L_{con}$ terms penalize differences among the subgraphs' adjacency matrices.

What would settle it

Train STMS-GCN with $L$, $K$, and $\lambda$ chosen on a held-out validation split and with several random seeds, then compare average MPJPE on the standard test portion of H3.6M and CMU Mocap against the same baselines; if the average no longer beats SPGSN and STBMP or the differences fall within run-to-run noise, the claimed superiority would be an artifact of test-set tuning.

Watch

Extended reading notes

Core claim

The central claim is that prediction quality improves when temporal and spatial information are modeled independently and then deliberately reconciled. The network has two orthogonal GCN branches: a temporal branch treats frames as nodes, and a spatial branch treats joints as nodes after a discrete cosine transform; a consistency loss $L_{ST}$ forces the intermediate predictions of the two branches to agree at every block layer, which the paper calls cross-domain knowledge transfer. Each branch uses $K$ trainable graph-convolution kernels whose outputs are averaged, and a homogeneous-information constraint $L_{con}$ keeps the adjacency matrices of the $K$ subgraphs similar so they do not diverge. The final prediction is the spatial branch output $\hat{Y}=Y_{S,L}$. Evaluated with MPJPE on H3.6M and CMU Mocap, the authors report that the full model outperforms the seven baselines in most cases, with best-average values 33.80 and 32.43 respectively.

Load-bearing premise

The reported margins assume that the hyperparameters $L=4$, $K=4$, and $\lambda=0.1$, selected on the same H3.6M benchmark, are not overfit to the test set; the paper reports no validation-set separation or repeated-seed variance.

Editorial extensions

If this is right

  • If the central claim is right, decoupled temporal and spatial GCN branches with a cross-domain consistency term are a sufficient recipe to beat the listed baselines on standard benchmarks, without new data or larger models.
  • The final-result choice matters: using the spatial branch output alone gives 33.80 average MPJPE on H3.6M, while averaging temporal and spatial outputs gives 33.98, so the paper's design decision contributes a small, measurable gain.
  • The reported gains are strongest at 80 ms (3.71%) and 160 ms (3.13%) on H3.6M and fade at 1000 ms, so the claim is best read as a short-to-mid-term forecasting improvement.
  • Regularizing adjacency matrices rather than weight matrices yields the best result (33.80 vs 33.98 for weights), indicating that consistency in graph structure is the more effective target.

Reading between the lines

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

  • Not in the paper: the consistency losses are a form of mutual regularization, so a direct test would be to give the temporal branch the same supervision from the spatial branch and see whether the dual-branch setup beats either single branch trained alone.
  • Not in the paper: the sharp performance drop at $\lambda=1.0$ (average 38.15) suggests the reported gains are sensitive to the constraint weight; a validation-set tuning protocol would be needed to know how much of the 3.71% improvement survives outside the benchmark.
  • Not in the paper: the same decouple-and-reconcile recipe could be tested on other sequence prediction tasks with graph structure, such as multi-agent trajectory forecasting, but the paper does not claim this.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes STMS-GCN, a graph convolutional architecture for 3D human motion prediction. The model runs two parallel GCN branches over temporal and spatial views of the input, averages K learned subgraph convolutions in each branch, and trains with three auxiliary consistency losses: a spatio-temporal loss that penalizes differences between the two branch predictions at every layer, and two subgraph-homogeneity losses that penalize differences between adjacency matrices. Experiments on H3.6M and CMU Mocap compare MPJPE with seven baselines; the authors report average improvements of 3.71% at 80 ms and 3.13% at 160 ms on H3.6M and an average MPJPE of 33.80 on H3.6M and 32.43 on CMU Mocap. Ablations show that each proposed component contributes to the selected configuration.

Significance. If the reported gains hold out of sample, the paper makes a modest but useful contribution: the dual-branch design with cross-domain consistency at multiple scales and the adjacency-matrix homogeneity regularizer are sensible, coherently described, and supported by ablations. The loss formulation is not circular, and the comparison covers standard benchmarks and relevant GCN baselines. However, the central claim is a small empirical margin, and the current evaluation does not establish that the margin generalizes: hyperparameters are selected on the same H3.6M test benchmark, no error bars or repeated-seed results are given, and some long-horizon margins are below 0.2 mm. The architectural ideas are worth publishing only if the evaluation protocol is strengthened.

major comments (4)
  1. [III-B (Effect of STMSBs (L) and graph convolutions (K); Influence of hyper-paramter λ)] The central claim of the paper is the numerical MPJPE improvement over baselines, but the hyperparameters L=4, K=4, and λ=0.1 are selected by best performance on H3.6M, the same benchmark used for the final comparison in Tables I and II. No validation split is described anywhere in Section III. Because the final table is the same evaluation used for hyperparameter selection, the reported margins (e.g., 3.71% at 80 ms and 3.13% at 160 ms) are optimistically biased and could be selection artifacts; this is load-bearing because the contribution is precisely that small empirical gain. The authors should either retune on a disjoint validation split and report test results only once, or report results across all hyperparameter settings and show that the chosen configuration is not a cherry-picked extremum.
  2. [Table I] No error bars, standard deviations, or repeated-seed results are reported. Several margins are within typical run-to-run variation: at 1000 ms on H3.6M the proposed method (109.51) is only 0.13 mm better than SPGSN (109.64) and 0.80 mm better than PGBIG (110.31), while at 560 ms PGBIG (76.91) is better than the proposed method (77.81). Without multiple seeds and a significance statement, the claim 'outperforms the baselines in most cases' is not supported at the long horizons, and even the short-horizon averages could be noise. Please report mean ± std over at least three seeds and, ideally, a paired test across actions.
  3. [Table III] The ablation study only evaluates the final selected configuration (L=4, K=4, λ=0.1) and does not re-tune hyperparameters for each ablated variant. Since removing a module changes the loss surface and the optimal λ, the contribution of each component is not isolated; the conclusion that 'all modules contribute' (Section III-B, 'Ablation of STMSB Block Design') is therefore weaker than claimed. Please report at least the sensitivity of each ablated variant to λ, or use a fixed validation-based protocol for all variants.
  4. [Section III-A and III-B] The experimental setup is not reproducible from the manuscript: there is no description of optimizer, learning rate, batch size, number of epochs, seeds, or the exact train/test split and preprocessing protocol (including how the 2D/3D D and padding are handled), and the external link in Fig. 3 only hosts videos, not code. Given that the reported improvements are small, the absence of these details makes independent verification impossible.
minor comments (6)
  1. [II-C, Eq. (7)] The quantity in Eq. (7) is described as an MPJPE constraint, but MPJPE is normally an error against ground truth; here it is the distance between the two branch predictions. Please rename it (e.g., cross-branch consistency loss) to avoid confusion.
  2. [Table I caption] The caption contains the typo 'Mothod'; it should read 'Method'.
  3. [Table IV caption] The caption contains 'Dfferent'; it should read 'Different'.
  4. [III-B, Influence of hyper-paramter λ] The text refers to 'Tab. 4' for the λ sweep, but the λ results appear in Fig. 4; Table IV concerns consistency constraints. Please correct the cross-reference.
  5. [Introduction] 'leaded' in Section I should be 'led'.
  6. [III-A] Please clarify what 'STBMP without incremental information' means and state whether the reported number is taken from the original STBMP paper or a reimplementation.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the empirical MPJPE comparison is self-contained; test-set hyperparameter selection is a generalization concern, not a definitional circularity.

full rationale

The paper's central claim is an empirical accuracy comparison on H3.6M and CMU-Mocap (Sec. III-B, Tables I-II), not a first-principles derivation. I checked each load-bearing step against the seven circularity patterns. The consistency losses (Eqs. 7, 9-10) and final loss (Eq. 12) are training objectives that constrain intermediate branch outputs; they do not define the reported MPJPE in terms of themselves. The final prediction is simply the spatial-branch decoder output YS,L (Eq. 6), and the multi-subgraph averaging (Eq. 8) is an architectural mechanism rather than a renamed empirical result. The only self-reference is the baseline STBMP [25], co-authored by the first author, but it is one of seven independently published baselines and is not used to justify the method's correctness; the stated 3.71% and 3.13% margins also depend on SPGSN and PGBIG numbers from the same tables. A legitimate concern is that L=4, K=4, and lambda=0.1 are chosen by best MPJPE on the same H3.6M benchmark used for the final evaluation (Sec. III-B, 'Effect of STMSBs (L) and graph convolutions (K)' and 'Influence of hyper-paramter λ'), with no validation split or repeated-seed evidence. However, that is a test-set-tuning / optimistic-bias issue for correctness and generalization, not a circular reduction: no equation collapses into its input, and no fitted parameter is relabeled as a predicted quantity. Apart from this evaluation-bias caveat, the derivation chain is self-contained and the comparison is externally anchored to standard benchmarks and published baselines.

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

The central claim is empirical. The main fitted quantities are loss weights and architecture sizes chosen by test-benchmark sweeps (lambda, beta, L, K). The learned weights W and adjacency matrices A are standard trainable parameters of the GCN and are not listed here. No new physical or conceptual entities are postulated; subgraphs, branches, and consistency losses are design choices inside the trained model.

free parameters (4)
  • lambda (loss weight) = 0.1
    Weight balancing L1 against consistency losses; best value found by grid search on the H3.6M benchmark (Fig. 4 / Tab. 4) and used for all reported results.
  • beta (homogeneity weight) = 0.1
    Weight for adjacency consistency terms, selected by test-benchmark sweep (Tab. V) and fixed throughout.
  • number of subgraphs K = 4
    Chosen by performance comparison in Fig. 3 rather than by theory; K=4 used for all reported results.
  • number of layers L = 4
    Selected experimentally from Fig. 3; L=4 used throughout the paper.
assumptions (4)
  • domain assumption MPJPE on H3.6M and CMU Mocap is a reliable quality metric for motion prediction.
    All conclusions about superiority rest on this loss; the paper gives no perceptual or downstream validation.
  • domain assumption The graph convolution update H = sigma(A X W) with trainable adjacency A is a valid model of kinematic dependencies.
    Standard assumption in GCN-based HMP, adopted in Sec. II-D without additional justification.
  • standard math DCT and IDCT are invertible enough to preserve pose information when followed by a learned decoder.
    DCT is orthonormal, so in the absence of learned corruption IDCT recovers the exact signal; the paper relies on this in Eq. (6).
  • domain assumption Padding the last observed pose by repetition is a valid input convention.
    The paper follows prior work [30], [31] in constructing X by repeating X_Tp, which biases the model toward predicting continuation of the last pose.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatio-Temporal Multi-Subgraph GCN for 3D Human Motion Prediction." pith.science (2026). https://pith.science/paper/D7T25USO

@misc{pith2026250100317,
  author       = {Pith},
  title        = {Pith review of: Spatio-Temporal Multi-Subgraph GCN for 3D Human Motion Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D7T25USO}},
  note         = {Machine review of arXiv:2501.00317}
}
read the original abstract

Human motion prediction (HMP) involves forecasting future human motion based on historical data. Graph Convolutional Networks (GCNs) have garnered widespread attention in this field for their proficiency in capturing relationships among joints in human motion. However, existing GCN-based methods tend to focus on either temporal-domain or spatial-domain features, or they combine spatio-temporal features without fully leveraging the complementarity and cross-dependency of these two features. In this paper, we propose the Spatial-Temporal Multi-Subgraph Graph Convolutional Network (STMS-GCN) to capture complex spatio-temporal dependencies in human motion. Specifically, we decouple the modeling of temporal and spatial dependencies, enabling cross-domain knowledge transfer at multiple scales through a spatio-temporal information consistency constraint mechanism. Besides, we utilize multiple subgraphs to extract richer motion information and enhance the learning associations of diverse subgraphs through a homogeneous information constraint mechanism. Extensive experiments on the standard HMP benchmarks demonstrate the superiority of our method.

Figures

Figures reproduced from arXiv: 2501.00317 by the authors.

Figure 1
Figure 1. Comparison of prediction methods. (a): Individual [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of STMS-GCN. where Fpred denotes a predictor, which bridges the past motion X to future motion Y. The aim of HMP is that the prediction Ye is as accurate as possible compared to Y. B. Spatio-Temporal Multi-Subgraph Block Current GCN-based methods suffer limitations by predom￾inantly focusing on either temporal or spatial features, disre￾garding the potential synergy between them [32], [33]. Despite some… view at source ↗
Figure 4
Figure 4. Impact of the fusion coefficient λ on the motion predictive performance. λ 80ms 160ms 320ms 400ms Avg. 0 9.69 22.06 48.44 60.24 35.11 0.001 9.80 21.98 47.63 59.23 34.66 0.01 9.61 21.75 47.26 58.65 34.32 0.1 9.61 21.63 46.40 57.55 33.80 1.0 10.83 24.71 52.58 64.46 38.15 method in the external link 1 . The results show that STMS￾GCN generates more precise and reasonable future poses. Ablation of STMSB Block Design. We… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 17 canonical work pages

  1. [25]

    Spatio- temporal branching for motion prediction using motion increments,

    J. Wang, Y . Zhou, W. Qiang, Y . Ba, B. Su, and J.-R. Wen, “Spatio- temporal branching for motion prediction using motion increments,” in Proceedings of the 31st ACM International Conference on Multimedia , 2023, pp. 4290–4299

  2. [1]

    Human-aware robotic assistant for collabo- rative assembly: Integrating human motion prediction with planning in time,

    V . V . Unhelkar, P. A. Lasota, Q. Tyroller, R.-D. Buhai, L. Marceau, B. Deml, and J. A. Shah, “Human-aware robotic assistant for collabo- rative assembly: Integrating human motion prediction with planning in time,” IEEE Robotics and Automation Letters , vol. 3, no. 3, pp. 2394– 2401, 2018

  3. [2]

    A neural temporal model for human motion prediction,

    A. Gopalakrishnan, A. Mali, D. Kifer, L. Giles, and A. G. Ororbia, “A neural temporal model for human motion prediction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 12 116–12 125

  4. [3]

    3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception,

    S. Chen, B. Liu, C. Feng, C. Vallespi-Gonzalez, and C. Wellington, “3d point cloud processing and learning for autonomous driving: Impacting map creation, localization, and perception,” IEEE Signal Processing Magazine, vol. 38, no. 1, pp. 68–86, 2020

  5. [4]

    Convolutional sequence to sequence model for human dynamics,

    C. Li, Z. Zhang, W. S. Lee, and G. H. Lee, “Convolutional sequence to sequence model for human dynamics,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 5226– 5234

  6. [5]

    Trajectorycnn: a new spatio-temporal feature learning network for human motion prediction,

    X. Liu, J. Yin, J. Liu, P. Ding, J. Liu, and H. Liu, “Trajectorycnn: a new spatio-temporal feature learning network for human motion prediction,” IEEE Transactions on Circuits and Systems for Video Technology , vol. 31, no. 6, pp. 2133–2146, 2020

  7. [6]

    Efficient human motion prediction using temporal convolutional generative adversarial network,

    Q. Cui, H. Sun, Y . Kong, X. Zhang, and Y . Li, “Efficient human motion prediction using temporal convolutional generative adversarial network,” Information Sciences, vol. 545, pp. 427–447, 2021

  8. [7]

    On human motion prediction using recurrent neural networks,

    J. Martinez, M. J. Black, and J. Romero, “On human motion prediction using recurrent neural networks,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 2891–2900

Show all 36 references
  1. [8]

    Context- aware human motion prediction,

    E. Corona, A. Pumarola, G. Alenya, and F. Moreno-Noguer, “Context- aware human motion prediction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 6992–7001

  2. [9]

    Long-term human motion prediction by modeling motion context and enhancing motion dynamic,

    Y . Tang, L. Ma, W. Liu, and W. Zheng, “Long-term human motion prediction by modeling motion context and enhancing motion dynamic,” arXiv preprint arXiv:1805.02513 , 2018

  3. [10]

    Video prediction recalling long-term motion context via memory alignment learning,

    S. Lee, H. G. Kim, D. H. Choi, H.-I. Kim, and Y . M. Ro, “Video prediction recalling long-term motion context via memory alignment learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 3054–3063

  4. [11]

    Learning progressive joint propagation for human motion prediction,

    Y . Cai, L. Huang, Y . Wang, T.-J. Cham, J. Cai, J. Yuan, J. Liu, X. Yang, Y . Zhu, X. Shenet al., “Learning progressive joint propagation for human motion prediction,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part V...

  5. [12]

    Pose trans- formers (potr): Human motion prediction with non-autoregressive trans- formers,

    A. Mart ´ınez-Gonz´alez, M. Villamizar, and J.-M. Odobez, “Pose trans- formers (potr): Human motion prediction with non-autoregressive trans- formers,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 2276–2284

  6. [13]

    Auxiliary tasks benefit 3d skeleton-based human motion prediction,

    C. Xu, R. T. Tan, Y . Tan, S. Chen, X. Wang, and Y . Wang, “Auxiliary tasks benefit 3d skeleton-based human motion prediction,” in Proceed- ings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 9509–9520

  7. [14]

    Skeleton- parted graph scattering networks for 3d human motion prediction,

    M. Li, S. Chen, Z. Zhang, L. Xie, Q. Tian, and Y . Zhang, “Skeleton- parted graph scattering networks for 3d human motion prediction,” in Computer Vision–ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part VI . Springer, 2022, pp. 18–36

  8. [15]

    Decompose more and aggregate better: Two closer looks at frequency representation learning for human motion prediction,

    X. Gao, S. Du, Y . Wu, and Y . Yang, “Decompose more and aggregate better: Two closer looks at frequency representation learning for human motion prediction,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2023, pp. 6451–6460

  9. [16]

    Dynamic composi- tional graph convolutional network for efficient composite human motion prediction,

    W. Zhang, S. Zhao, F. Meng, S. Wu, and M. Liu, “Dynamic composi- tional graph convolutional network for efficient composite human motion prediction,” in Proceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 2856–2864

  10. [17]

    Skeleton- based online action prediction using scale selection network,

    J. Liu, A. Shahroudy, G. Wang, L.-Y . Duan, and A. C. Kot, “Skeleton- based online action prediction using scale selection network,” IEEE transactions on pattern analysis and machine intelligence, vol. 42, no. 6, pp. 1453–1467, 2019

  11. [18]

    Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction,

    M. Li, S. Chen, Y . Zhao, Y . Zhang, Y . Wang, and Q. Tian, “Dynamic multiscale graph neural networks for 3d skeleton based human motion prediction,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 214–223

  12. [19]

    Msr-gcn: Multi-scale residual graph convolution networks for human motion prediction,

    L. Dang, Y . Nie, C. Long, Q. Zhang, and G. Li, “Msr-gcn: Multi-scale residual graph convolution networks for human motion prediction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 11 467–11 476

  13. [20]

    Towards accurate 3d human motion prediction from incomplete observations,

    Q. Cui and H. Sun, “Towards accurate 3d human motion prediction from incomplete observations,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 4801–4810

  14. [21]

    Progressively generating better initial guesses towards next stages for high-quality human motion prediction,

    T. Ma, Y . Nie, C. Long, Q. Zhang, and G. Li, “Progressively generating better initial guesses towards next stages for high-quality human motion prediction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 6437–6446

  15. [22]

    Masked autoencoders as spatiotemporal learners,

    C. Feichtenhofer, Y . Li, K. He et al. , “Masked autoencoders as spatiotemporal learners,” Advances in neural information processing systems, vol. 35, pp. 35 946–35 958, 2022

  16. [23]

    Temporal attention unit: Towards efficient spatiotemporal predictive learning,

    C. Tan, Z. Gao, L. Wu, Y . Xu, J. Xia, S. Li, and S. Z. Li, “Temporal attention unit: Towards efficient spatiotemporal predictive learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 18 770–18 782

  17. [24]

    Spatio-temporal gating-adjacency gcn for human motion prediction. 2022 ieee,

    C. Zhong, L. Hu, Z. Zhang, Y . Ye, and S. hong Xia, “Spatio-temporal gating-adjacency gcn for human motion prediction. 2022 ieee,” in CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022, pp. 6437–6446

  18. [26]

    Learning factored represen- tations in a deep mixture of experts,

    D. Eigen, M. Ranzato, and I. Sutskever, “Learning factored represen- tations in a deep mixture of experts,” arXiv preprint arXiv:1312.4314 , 2013

  19. [27]

    Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,

    N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” arXiv preprint arXiv:1701.06538 , 2017

  20. [28]

    Diversifying the mixture-of-experts representation for language models with orthogonal optimizer,

    B. Liu, L. Ding, L. Shen, K. Peng, Y . Cao, D. Cheng, and D. Tao, “Diversifying the mixture-of-experts representation for language models with orthogonal optimizer,” arXiv preprint arXiv:2310.09762 , 2023

  21. [29]

    A closer look into mixture-of-experts in large language models,

    K. M. Lo, Z. Huang, Z. Qiu, Z. Wang, and J. Fu, “A closer look into mixture-of-experts in large language models,” arXiv preprint arXiv:2406.18219, 2024

  22. [30]

    Learning trajectory depen- dencies for human motion prediction,

    W. Mao, M. Liu, M. Salzmann, and H. Li, “Learning trajectory depen- dencies for human motion prediction,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2019, pp. 9489–9497

  23. [31]

    History repeats itself: Human motion prediction via motion attention,

    W. Mao, M. Liu, and M. Salzmann, “History repeats itself: Human motion prediction via motion attention,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XIV 16 . Springer, 2020, pp. 474–489

  24. [32]

    Fast and furious: Real time end- to-end 3d detection, tracking and motion forecasting with a single convolutional net,

    W. Luo, B. Yang, and R. Urtasun, “Fast and furious: Real time end- to-end 3d detection, tracking and motion forecasting with a single convolutional net,” in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition , 2018, pp. 3569–3577

  25. [33]

    Skele- ton graph scattering networks for 3d skeleton-based human motion prediction,

    M. Li, S. Chen, Z. Liu, Z. Zhang, L. Xie, Q. Tian, and Y . Zhang, “Skele- ton graph scattering networks for 3d skeleton-based human motion prediction,” in Proceedings of the IEEE/CVF international conference on computer vision , 2021, pp. 854–864

  26. [34]

    Human motion predic- tion via spatio-temporal inpainting,

    A. Hernandez, J. Gall, and F. Moreno-Noguer, “Human motion predic- tion via spatio-temporal inpainting,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2019, pp. 7134–7143

  27. [35]

    Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments,

    C. Ionescu, D. Papava, V . Olaru, and C. Sminchisescu, “Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments,” IEEE transactions on pattern analysis and machine intelligence, vol. 36, no. 7, pp. 1325–1339, 2013

  28. [36]

    Space-time- separable graph convolutional network for pose forecasting,

    T. Sofianos, A. Sampieri, L. Franco, and F. Galasso, “Space-time- separable graph convolutional network for pose forecasting,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 11 209–11 218

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.