Pith. sign in

REVIEW 3 major objections 5 minor 50 references

Efficient Multi-Person Motion Prediction by Lightweight Spatial and Temporal Interactions

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims a 0.04M-parameter network of linear temporal updates and cross-level affine interactions outperforms transformer-based multi-person motion predictors on standard benchmarks.

desk verdict Credible lightweight MLP-based multi-person motion predictor with solid 3DPW results, but the broad SOTA claim over synthetic benchmarks needs the generation script and error bars before I'd trust it. read the letter →

arxiv 2507.09446 v1 pith:YJMTEUYJ submitted 2025-07-13 cs.CV

Yuanhong Zheng , Ruixuan Yu , Jian Sun This is my paper
classification cs.CV
keywords 3Dmulti-personmotionpredictionlightweightneuralnetworkspatio-temporalinteractionMLP-basedarchitectureinter-persondistanceembeddingpermutationinvariancecross-levelposeforecasting
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 sets out to show that 3D multi-person motion prediction does not need heavy Transformer stacks. Its EMPMP network uses two lightweight branches, one for each person's own motion and one for the whole scene, and lets them exchange information through learned affine transformations rather than attention. On 3DPW, CMU-Mocap, and MuPoTS-3D, the authors report lower MPJPE, VIM, and APE than prior methods while using as little as 0.04M parameters. The payoff, if true, is that accurate multi-person interaction modeling can run at a fraction of the computational cost and with a much simpler architecture.

What carries the argument

The Multi-level Estimation (ME) block is a dual-stream module: a local stream applies shared linear transformations along the time dimension for each person, and a global stream applies linear transformations along the merged person-time dimension to model the whole scene. The Cross-level Interaction (CI) block refines local features by affine scale-and-shift transformations generated from global features, and refines global features by a translation generated from local features, with an inter-person distance embedding computed from pairwise hip distances fed through a linear layer. Discrete Cosine Transform compresses the temporal input into low-frequency coefficients, and Permutation-Invariant Person Sorting reorders people by descending pairwise distance so predictions are unchanged by person order.

What would settle it

Run EMPMP and the same baselines on the original multi-range-transformer synthetic CMU-Syn splits, or on an independently re-created version of that exact protocol, and check whether EMPMP still leads on MPJPE, VIM, and APE; if its margin disappears or reverses, the state-of-the-art claim on CMU-Mocap and MuPoTS-3D depends on the authors' synthesis rather than on the method.

Watch

Extended reading notes

Core claim

EMPMP achieves state-of-the-art accuracy on standard multi-person motion prediction benchmarks with 0.04M to 0.65M parameters and 1.67G FLOPs, by replacing attention with iterated linear temporal updates and cross-level affine interactions, augmented by an explicit inter-person hip-distance embedding. The authors report this across 3DPW-Ori, 3DPW-RC, AMASS/3DPW, CMU-Syn, and CMU-Syn/MuPoTS-3D settings, with the clearest margin on 3DPW-Ori where MPJPE drops to 131.8 from 137.5 for the best prior method and APE drops to 98.6 from 115.1. The claim is that the dual-branch local-global design plus bidirectional cross-level refinement captures both individual dynamics and social interactions without attention's parameter cost.

Load-bearing premise

The CMU-Mocap and MuPoTS-3D comparisons rest on the authors' own re-synthesized three-person scenes, since the original synthesized datasets are not public; if those scenes differ from the original protocol, the reported state-of-the-art results on those benchmarks would not carry over.

Editorial extensions

If this is right

  • If the reported results hold, interaction modeling in multi-person motion prediction can be done without attention: linear temporal mixing, cross-level affine gating, and a distance signal are sufficient.
  • At 0.04M parameters and 1.67G FLOPs, the model fits on-device budgets, making real-time multi-person forecasting feasible on mobile or embedded hardware.
  • The 0.65M variant with added spatial updates shows the architecture scales with capacity, so larger versions of the same design may push accuracy higher while remaining far below transformer parameter counts.
  • DCT-based low-frequency input and permutation-invariant person sorting make the model robust to person ordering and camera displacement, simplifying deployment in unconstrained scenes.

Reading between the lines

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

  • Editorial: the same local-linear/global-linear update plus cross-level affine gate recipe could transfer to other multi-agent sequence tasks, such as point-agent trajectory prediction, where interaction modeling currently relies on heavy attention; this is not claimed in the paper.
  • Editorial: the inter-person distance embedding is a cheap inductive bias; a testable extension is to replace hip distance with learned pairwise joint-pair distances or collision-related features and check whether accuracy improves further.
  • Editorial: because the authors had to re-synthesize the CMU-Mocap protocol, an independent comparison on the original released splits, when they become available, would be the cleanest test of whether the SOTA claim generalizes beyond the authors' own synthesis distribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. EMPMP is an MLP-based architecture for multi-person 3D motion prediction. It starts with a Permutation-Invariant Person Sorting (PIPS) stage to canonicalize person order, applies DCT along time, embeds joints, and then alternates Multi-level Estimation (ME) blocks (local per-person temporal linear updates and a global person-temporal linear update) with Cross-level Interaction (CI) blocks that refine local and global representations through learned affine transformations and an inter-person hip-distance embedding. Future motion is produced by inverse transforms, and training uses a joint-position plus velocity loss. The paper reports state-of-the-art MPJPE/VIM/APE results on 3DPW-Ori, 3DPW-RC, on a self-generated CMU-Syn benchmark, and on CMU-Syn/MuPoTS cross-dataset evaluation, with parameter counts of 0.04M to 0.65M and substantially lower FLOPs than transformer baselines. Ablations isolate the contributions of the local/global updates, the cross-level refinements, the distance embedding, and the loss components.

Significance. If the reported results hold, the paper makes a useful contribution: it is one of the first MLP-only multi-person motion predictors, it demonstrates a dramatic parameter and FLOP reduction relative to transformer baselines, and the 3DPW experiments follow a public protocol with publicly available data. The architectural ideas—dual local/global streams, affine cross-level interaction, and an inter-person distance embedding—are clearly described and are supported by a reasonably complete set of ablations. The main reservations concern verification and a specific technical flaw in the claimed permutation invariance: the headline SOTA claim on CMU-Mocap and MuPoTS-3D rests on privately regenerated synthetic benchmarks, and the PIPS sorting rule degenerates for the two-person scenes used in 3DPW. These issues must be resolved before the broad SOTA claim can be accepted; the 3DPW results alone are promising but do not by themselves support the full abstract claim.

major comments (3)
  1. [Sec. 4.1; Supplementary A.1] The CMU-Syn and CMU-Syn/MuPoTS results in Tables 1 and 3 are produced on synthetic scenes regenerated by the authors, because "synthesized datasets from prior works are not publicly available" (Supp. A.1). The paper states only that the data were generated "following approach in [46]"; it does not provide the generation script, the random splits, the scene-composition rules, the hip-joint scaling, or the interaction-pattern sampling. Since the official code of the compared methods was run on this private regeneration, the internal comparisons may be self-consistent, but the resulting numbers are not comparable to any public benchmark. The abstract's claim of state-of-the-art performance on "standard datasets of CMU-Mocap, MuPoTS-3D, and 3DPW" therefore rests on unverifiable data for two of the three named datasets. Please release the synthesis code/data or restrict the SOTA claim to 3DPW. In addition, no error bars or statistical significance tests are reported; given the random 30-frame sampling in the 1s/1s CMU-Syn setting, a single run may not support fine-grained rankings.
  2. [Supplementary B, Eq. (B.1); Sec. 3.1] The PIPS sorting rule does not provide a canonical permutation for the two-person scenes used in 3DPW. For P=2, dp_j is the same for both persons because each person has the same distance to the other; a descending sort of equal values is not a canonical ordering and, with a stable sort, leaves the input order unchanged. The global temporal linear transform (Eq. 3) and the distance embedding (Eq. 7) are not permutation-equivariant, so without a valid canonical order the model is not actually order-agnostic and the "permutation-invariant prediction" claim in Sec. 3.1 and Supplementary B is unsupported. Please specify a deterministic tie-breaking rule and provide an empirical test—for example, predictions under all person permutations for both P=2 and P=3 scenes—or replace PIPS with a permutation-equivariant formulation.
  3. [Table 3; Sec. 4.2] The 1s/3s CMU-Syn results in Table 3 are unreadable as typeset: entries such as "89 145 18965 97 12886 102 111" do not separate the nine metric-by-horizon columns, and similar issues affect the compressed tables in the supplementary material. Because the 1s/3s extended-forecasting claim is presented as a separate result, the table must be reformatted so that each reported number can be assigned unambiguously to a metric and an output length.
minor comments (5)
  1. [Abstract vs. Sec. 3.5] The abstract and GitHub URL state that code is available, while Sec. 3.5 says "Our code will be released if the paper is accepted." Please reconcile these statements; in particular, the synthetic-data generation scripts should be released if the synthetic-benchmark SOTA claim is retained.
  2. [Table 1 header] Table 1's header is ambiguous: the CMU-Syn and CMU-Syn/MuPoTS blocks each contain two settings (2s/2s and 1s/1s), but the "In/Out Length" row does not clearly list the eight column lengths. Adding explicit subheadings for each setting would avoid confusion about which numbers correspond to which benchmark.
  3. [Sec. 4.1; Supplementary A.2] The main text describes VIM as "mean joint-coordinate distance," but Eq. (A.2) computes a per-person root-sum-of-squares error averaged over persons. Please align the prose with the formula or revise the formula.
  4. [Supplementary A.2, Eq. (A.5)] The FDE definition averages over J joints even though the hip displacement is identical for every joint; this over-counts the joint index and should be simplified to an average over persons only (or the formula should be clarified).
  5. [Sec. 3.1 vs. Supplementary B] The main text says PIPS reorders individuals "by descending pairwise distances," whereas Eq. (B.1) sorts by the sum of distances from each person to all others. Please use consistent wording so the actual criterion is unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the derivation is self-contained and the synthetic-benchmark regeneration is a reproducibility gap, not a circular reduction.

full rationale

The paper's derivation chain is entirely architectural and empirical: the model takes observed multi-person joint histories X, applies permutation-invariant sorting (PIPS), DCT, linear embeddings, iterative local/global temporal updates, affine cross-level interactions, an inter-person distance embedding computed from observed hip inputs, and a linear prediction head, with a standard mean-joint plus velocity loss. None of these components are defined in terms of the target predictions or the reported metrics. The distance embedding in Eq. (6) uses only observed historical hip positions, not ground-truth future frames, so there is no leakage or self-definition. The hyperparameters (alpha, N, M, K, C) are selected by ablation studies on the evaluation benchmarks, which is standard practice and does not amount to fitting a parameter and then renaming it a prediction. The paper contains no self-citations by the present authors, and no uniqueness theorem or prior-work ansatz is invoked to force the architecture. The only load-bearing concern is that the CMU-Syn and CMU-Syn/MuPoTS benchmarks were regenerated by the authors because 'synthesized datasets from prior works are not publicly available' (Supplementary A.1), so those two columns of Table 1 cannot be independently verified against the original MRT protocol; however, the paper states that official codes of compared methods were run on the same synthesized versions, making the internal comparison self-consistent. That is a validity and reproducibility limitation, not a circular reduction of the kind where an equation or fitted value is equivalent to its own input. A minor availability contradiction also exists between the abstract's 'Code is available' and Section 3.5's 'Our code will be released if the paper is accepted,' which is a presentation flaw rather than circularity. Overall, the core claims rest on external benchmarks (3DPW) plus a transparently described regeneration protocol, and no derivation step reduces to its own inputs.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

No new physical entities or conserved quantities are introduced; the modules (PIPS, ME, CI) are algorithmic operations. The free parameters are architectural capacity choices and fusion weights selected by ablations. The axioms are the unverified corners of the benchmark reproduction and the sufficiency of hip-distance as the interaction signal.

free parameters (6)
  • combination parameter alpha = 0.2
    Eq. (11) fusion weight between local and global features; tuned on 3DPW-RC (Tab. D.3).
  • local update iterations N = 16
    Number of T-linear iterations in ME block; sweep in Fig. 4a.
  • global update iterations M = 1
    Number of PT-linear iterations; sweep in Fig. 4b; M<=2 preferred to avoid oversmoothing.
  • ESTFL stages K = 4
    Stack depth of ME+CI blocks; sweep in Fig. 4c.
  • feature channels C = 45 (CMU/MuPoTS), 39 (3DPW)
    Embedding dimension chosen per dataset; hand-set capacity.
  • pre-training model capacity = 0.65M params
    A larger variant is used for AMASS/3DPW settings to ensure fair comparison (Supp. B); chosen by the authors.
assumptions (4)
  • domain assumption The authors' synthesized CMU-Syn and CMU-Syn/MuPoTS data faithfully reproduce the MRT protocol.
    Supp. A.1 states the previous synthetic data is unavailable and all experiments were conducted on the authors' synthesized versions; the comparability of every CMU-Mocap/MuPoTS baseline depends on this.
  • domain assumption The hip-to-hip Euclidean distance is a sufficient signal for inter-person spatial interaction.
    Eq. (6) defines the distance matrix using only hip joints; the method claims this captures the spatial interactions between individuals.
  • ad hoc to paper PIPS produces a deterministic and meaningful person ordering for any number of persons.
    Eq. (B.1) sorts by the sum of pairwise first-frame hip distances; for P=2 both persons have equal sums, so the order is degenerate unless an unspecified tie-break exists.
  • domain assumption The fixed linear layer over the flattened person-time dimension (W_g in Eq. 3) can capture cross-person temporal dependencies.
    The global stream models interaction as a single linear map over PT; nonlinear interactions must be captured implicitly through LayerNorm and the CI affine modulations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Multi-Person Motion Prediction by Lightweight Spatial and Temporal Interactions." pith.science (2026). https://pith.science/paper/YJMTEUYJ

@misc{pith2026250709446,
  author       = {Pith},
  title        = {Pith review of: Efficient Multi-Person Motion Prediction by Lightweight Spatial and Temporal Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJMTEUYJ}},
  note         = {Machine review of arXiv:2507.09446}
}
read the original abstract

3D multi-person motion prediction is a highly complex task, primarily due to the dependencies on both individual past movements and the interactions between agents. Moreover, effectively modeling these interactions often incurs substantial computational costs. In this work, we propose a computationally efficient model for multi-person motion prediction by simplifying spatial and temporal interactions. Our approach begins with the design of lightweight dual branches that learn local and global representations for individual and multiple persons separately. Additionally, we introduce a novel cross-level interaction block to integrate the spatial and temporal representations from both branches. To further enhance interaction modeling, we explicitly incorporate the spatial inter-person distance embedding. With above efficient temporal and spatial design, we achieve state-of-the-art performance for multiple metrics on standard datasets of CMU-Mocap, MuPoTS-3D, and 3DPW, while significantly reducing the computational cost. Code is available at https://github.com/Yuanhong-Zheng/EMPMP.

Figures

Figures reproduced from arXiv: 2507.09446 by the authors.

Figure 1
Figure 1. Comparison of our model with state-of-the-art methods [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Framework of our EMPMP network. The input is mapped to order-agnostic feature space by PIPS and DCT stages. The Joint [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Qualitative results in the 3DPW-RC setting involve predicting 14 frames using 16 frames of input. Different colors are used to distinguish between individuals in the sequence. 4 8 12 16 20 24 28 N 128 130 132 134 136 MPJPE (a) 1 2 4 6 8 10 M 128 130 132 134 136 138 (b) 1 2 3 4 5 6 K 128 130 132 134 136 138 (c) CMU-Syn 3DPW-Ori SOTA(3DPW-Ori) SOTA(CMU-Syn) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance curves with varied hyper-parameters in [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 45 canonical work pages

  1. [46]

    Multi-person 3d motion prediction with multi- range transformers

    Jiashun Wang, Huazhe Xu, Medhini Narasimhan, and Xiao- long Wang. Multi-person 3d motion prediction with multi- range transformers. Advances in Neural Information Pro- cessing Systems, 34:6036–6049, 2021. 1, 2, 5, 6, 7, 4

    work page 2021

  2. [1]

    http:// mocap.cs.cmu.edu/

    Cmu graphics lab motion capture database. http:// mocap.cs.cmu.edu/. 5

    work page

  3. [2]

    Socially and contextually aware human motion and pose forecasting

    Vida Adeli, Ehsan Adeli, Ian Reid, Juan Carlos Niebles, and Hamid Rezatofighi. Socially and contextually aware human motion and pose forecasting. IEEE Robotics and Automation Letters, 5(4):6033–6040, 2020. 1

    work page 2020

  4. [3]

    Tripod: Human trajectory and pose dynamics forecasting in the wild

    Vida Adeli, Mahsa Ehsanpour, Ian Reid, Juan Car- los Niebles, Silvio Savarese, Ehsan Adeli, and Hamid Rezatofighi. Tripod: Human trajectory and pose dynamics forecasting in the wild. In Proceedings of the IEEE/CVF In- ternational Conference on Computer Vision , pages 13390– 13400, 2021. 6, 1

    work page 2021

  5. [4]

    Discrete cosine transform

    Nasir Ahmed, T Natarajan, and Kamisetty Ramamohan Rao. Discrete cosine transform. IEEE Transactions on Comput- ers, 100(1):90–93, 1974. 2

    work page 1974

  6. [5]

    Latentformer: Multi-agent transformer- based interaction modeling and trajectory prediction

    Elmira Amirloo, Amir Rasouli, Peter Lakner, Mohsen Ro- hani, and Jun Luo. Latentformer: Multi-agent transformer- based interaction modeling and trajectory prediction. arXiv preprint arXiv:2203.01880, 2022. 1

    arXiv 2022

  7. [6]

    Systematic review of multimodal human– computer interaction

    Jose Daniel Azofeifa, Julieta Noguez, Sergio Ruiz, Jos´e Mart ´ın Molina-Espinosa, Alejandra J Magana, and Bedrich Benes. Systematic review of multimodal human– computer interaction. Informatics, 9(1):13, 2022. 1

    work page 2022

  8. [7]

    A sys- tematic review of the application of camera-based human pose estimation in the field of sport and physical exercise

    Aritz Badiola-Bengoa and Amaia Mendez-Zorrilla. A sys- tematic review of the application of camera-based human pose estimation in the field of sport and physical exercise. Sensors, 21(18):5996, 2021. 1

    work page 2021

Show all 50 references
  1. [8]

    Motionmixer: Mlp- based 3d human body pose forecasting

    Arij Bouazizi, Adrian Holzbock, Ulrich Kressel, Klaus Di- etmayer, and Vasileios Belagiannis. Motionmixer: Mlp- based 3d human body pose forecasting. arXiv preprint arXiv:2207.00499, 2022. 2

    2022 arXiv

  2. [9]

    Surveillance systems evaluation: a systematic review of the existing approaches

    Clementine Calba, Flavie L Goutard, Linda Hoinville, Pas- cal Hendrikx, Ann Lindberg, Claude Saegerman, and Marisa Peyre. Surveillance systems evaluation: a systematic review of the existing approaches. BMC Public Health , 15:1–13,

  3. [10]

    Cyclemlp: A mlp-like architecture for dense prediction

    Shoufa Chen, Enze Xie, Chongjian Ge, Runjian Chen, Ding Liang, and Ping Luo. Cyclemlp: A mlp-like architecture for dense prediction. arXiv preprint arXiv:2107.10224, 2021. 2

    2021 arXiv

  4. [11]

    Vnagt: Variational non-autoregressive graph transformer network for multi-agent trajectory prediction

    Xiaobo Chen, Huanjia Zhang, Yu Hu, Jun Liang, and Hai Wang. Vnagt: Variational non-autoregressive graph transformer network for multi-agent trajectory prediction. IEEE Transactions on Vehicular Technology, 72(10):12540– 12552, 2023. 1

    2023

  5. [12]

    Generating long sequences with sparse transformers

    Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 2

    1904 arXiv

  6. [13]

    Action-agnostic human pose fore- casting

    Hsu-kuang Chiu, Ehsan Adeli, Borui Wang, De-An Huang, and Juan Carlos Niebles. Action-agnostic human pose fore- casting. In Proceedings of the IEEE/CVF Winter Confer- ence on Applications of Computer Vision, pages 1423–1432,

  7. [14]

    A system for pose analysis and selection in virtual reality environments

    Andrew Clark, Anban W Pillay, and Deshendran Moodley. A system for pose analysis and selection in virtual reality environments. In Conference of the South African Institute of Computer Scientists and Information Technologists, pages 210–216, 2020. 1

    2020

  8. [15]

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

    Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Msr-gcn: Multi-scale residual graph convolution networks for human motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11467–11476, 2021. 2

    2021

  9. [16]

    Recurrent network models for human dynam- ics

    Katerina Fragkiadaki, Sergey Levine, Panna Felsen, and Ji- tendra Malik. Recurrent network models for human dynam- ics. In Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, pages 4346–4354, 2015. 1, 2

    2015

  10. [17]

    Capture of 3d human motion pose in virtual reality based on video recognition

    Qiang Fu, Xingui Zhang, Jinxiu Xu, and Haimin Zhang. Capture of 3d human motion pose in virtual reality based on video recognition. Complexity, 2020(1):8857748, 2020. 1

    2020

  11. [18]

    Back to mlp: A simple baseline for human motion prediction

    Wen Guo, Yuming Du, Xi Shen, Vincent Lepetit, Xavier Alameda-Pineda, and Francesc Moreno-Noguer. Back to mlp: A simple baseline for human motion prediction. In Proceedings of the IEEE/CVF Winter Conference on Appli- cations of Computer Vision, pages 4809–4819, 2023. 2, 5

    2023

  12. [19]

    Multi-shot human re-identification for the security in video surveillance systems

    Bassem HadjKacem, Walid Ayedi, Mohamed Abid, and Hichem Snoussi. Multi-shot human re-identification for the security in video surveillance systems. InIEEE 27th Interna- tional Conference on Enabling Technologies: Infrastructure for Collaborative Enterprises, pages 203–208, 2018. 1

    2018

  13. [20]

    Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3. 6m: Large scale datasets and pre- dictive methods for 3d human sensing in natural environ- ments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325–1339, 2013. 6

    2013

  14. [21]

    Multi-agent long-term 3d human pose forecasting via interaction-aware trajectory conditioning

    Jaewoo Jeong, Daehee Park, and Kuk-Jin Yoon. Multi-agent long-term 3d human pose forecasting via interaction-aware trajectory conditioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 1617–1628, 2024. 1, 2, 6, 7, 4, 5

    2024

  15. [22]

    Mart: Multiscale relational transformer net- works for multi-agent trajectory prediction

    Seongju Lee, Junseok Lee, Yeonguk Yu, Taeri Kim, and Kyoobin Lee. Mart: Multiscale relational transformer net- works for multi-agent trajectory prediction. In Proceedings of the European conference on computer vision , pages 89– 107, 2025. 1

    2025

  16. [23]

    Efficient nonlinear markov models for human mo- tion

    Andreas M Lehrmann, Peter V Gehler, and Sebastian Nowozin. Efficient nonlinear markov models for human mo- tion. In Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition , pages 1314–1321,

  17. [24]

    Graphmlp: A graph mlp-like ar- chitecture for 3d human pose estimation

    Wenhao Li, Mengyuan Liu, Hong Liu, Tianyu Guo, Ti Wang, Hao Tang, and Nicu Sebe. Graphmlp: A graph mlp-like ar- chitecture for 3d human pose estimation. Pattern Recogni- tion, 158:110925, 2025. 2

    2025

  18. [25]

    Baseball swing pose estimation using openpose

    Yungche Li, Chingtang Chang, Chinchang Cheng, and Yulen Huang. Baseball swing pose estimation using openpose. In IEEE International Conference on Robotics, Automation and Artificial Intelligence, pages 6–9, 2021. 1

    2021

  19. [26]

    Arhpe: Asymmetric relation- aware representation learning for head pose estimation in in- dustrial human–computer interaction

    Hai Liu, Tingting Liu, Zhaoli Zhang, Arun Kumar Sanga- iah, Bing Yang, and Youfu Li. Arhpe: Asymmetric relation- aware representation learning for head pose estimation in in- dustrial human–computer interaction. IEEE Transactions on Industrial Informatics, 18(10):7107–7117, 2022. 1

    2022

  20. [27]

    Transkd: Transformer knowledge distilla- tion for efficient semantic segmentation

    Ruiping Liu, Kailun Yang, Alina Roitberg, Jiaming Zhang, Kunyu Peng, Huayao Liu, Yaonan Wang, and Rainer Stiefelhagen. Transkd: Transformer knowledge distilla- tion for efficient semantic segmentation. arXiv preprint arXiv:2202.13393, 2022. 2

    2022 arXiv

  21. [28]

    Towards natural and accurate future motion prediction of humans and animals

    Zhenguang Liu, Shuang Wu, Shuyuan Jin, Qi Liu, Shijian Lu, Roger Zimmermann, and Li Cheng. Towards natural and accurate future motion prediction of humans and animals. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10004–10012, 2019. 1, 2

    2019

  22. [29]

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

    Tiezheng Ma, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing 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 , pages 643...

    2022

  23. [30]

    Amass: Archive of motion capture as surface shapes

    Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Ger- ard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 5442–5451, 2019. 5

    2019

  24. [31]

    Learning trajectory dependencies for human motion pre- diction

    Wei Mao, Miaomiao Liu, Mathieu Salzmann, and Hongdong Li. Learning trajectory dependencies for human motion pre- diction. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 9489–9497, 2019. 1, 2

    2019

  25. [32]

    History re- peats itself: Human motion prediction via motion attention

    Wei Mao, Miaomiao Liu, and Mathieu Salzmann. History re- peats itself: Human motion prediction via motion attention. In Proceedings of the European conference on computer vi- sion, pages 474–489, 2020. 1, 2

    2020

  26. [33]

    Single-shot multi-person 3d pose estimation from monocular rgb

    Dushyant Mehta, Oleksandr Sotnychenko, Franziska Mueller, Weipeng Xu, Srinath Sridhar, Gerard Pons-Moll, and Christian Theobalt. Single-shot multi-person 3d pose estimation from monocular rgb. In IEEE International Conference on 3D Vision, pages 120–130, 2018. 5

    2018

  27. [34]

    Trajectory- aware body interaction transformer for multi-person pose forecasting

    Xiaogang Peng, Siyuan Mao, and Zizhao Wu. Trajectory- aware body interaction transformer for multi-person pose forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17121– 17130, 2023. 1, 2, 6, 7, 4, 5

    2023

  28. [35]

    Vision based hand gesture recognition for human computer interaction: a survey

    Siddharth S Rautaray and Anupam Agrawal. Vision based hand gesture recognition for human computer interaction: a survey. Artificial Intelligence Review, 43:1–54, 2015. 1

    2015

  29. [36]

    Human-engaged computing: the future of human–computer interaction

    Xiangshi Ren, Chaklam Silpasuwanchai, and John Cahill. Human-engaged computing: the future of human–computer interaction. CCF Transactions on Pervasive Computing and Interaction, 1:47–68, 2019. 1

    2019

  30. [37]

    Display methods of projection augmented reality based on deep learning pose estimation

    Hyocheol Ro, Yoon Jung Park, Jung-Hyun Byun, and Tack- Don Han. Display methods of projection augmented reality based on deep learning pose estimation. InACM SIGGRAPH Posters, pages 1–2. 2019. 1

    2019

  31. [38]

    Immersion of virtual reality for rehabilitation-review.Applied Ergonomics, 69:153–161, 2018

    Tyler Rose, Chang S Nam, and Karen B Chen. Immersion of virtual reality for rehabilitation-review.Applied Ergonomics, 69:153–161, 2018. 1

    2018

  32. [39]

    Layer-wise pruning of transformer attention heads for efficient language modeling

    Kyuhong Shim, Iksoo Choi, Wonyong Sung, and Jungwook Choi. Layer-wise pruning of transformer attention heads for efficient language modeling. InIEEE 18th International SoC Design Conference, pages 357–358, 2021. 2

    2021

  33. [40]

    3d human pose estimation with spatio- temporal criss-cross attention

    Zhenhua Tang, Zhaofan Qiu, Yanbin Hao, Richang Hong, and Ting Yao. 3d human pose estimation with spatio- temporal criss-cross attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4790–4799, 2023. 1

    2023

  34. [41]

    Mlp-mixer: An all-mlp architecture for vision

    Ilya O Tolstikhin, Neil Houlsby, Alexander Kolesnikov, Lu- cas Beyer, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, et al. Mlp-mixer: An all-mlp architecture for vision. Advances in Neural Information Processing Systems, 34:2...

  35. [42]

    Somoformer: Multi-person pose forecasting with transformers

    Edward Vendrow, Satyajit Kumar, Ehsan Adeli, and Hamid Rezatofighi. Somoformer: Multi-person pose forecasting with transformers. arXiv preprint arXiv:2208.14023, 2022. 1, 2, 5, 6, 7, 4

    2022 arXiv

  36. [43]

    Recovering ac- curate 3d human pose in the wild using imus and a moving camera

    Timo V on Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering ac- curate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European conference on com- puter vision, pages 601–617, 2018. 5

    2018

  37. [44]

    Pvred: A position-velocity recurrent encoder-decoder for hu- man motion prediction

    Hongsong Wang, Jian Dong, Bin Cheng, and Jiashi Feng. Pvred: A position-velocity recurrent encoder-decoder for hu- man motion prediction. IEEE Transactions on Image Pro- cessing, 30:6096–6106, 2021. 2

    2021

  38. [45]

    Gaussian process dynamical models

    Jack Wang, Aaron Hertzmann, and David J Fleet. Gaussian process dynamical models. Advances in Neural Information Processing Systems, 18, 2005. 2

    2005

  39. [47]

    Eqmo- tion: Equivariant multi-agent motion prediction with invari- ant interaction reasoning

    Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmo- tion: Equivariant multi-agent motion prediction with invari- ant interaction reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition ,...

    2023

  40. [48]

    Joint- relation transformer for multi-person motion prediction

    Qingyao Xu, Weibo Mao, Jingze Gong, Chenxin Xu, Si- heng Chen, Weidi Xie, Ya Zhang, and Yanfeng Wang. Joint- relation transformer for multi-person motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9816–9826, 2023. 1, 2, 5, 6, 7, 4

    2023

  41. [49]

    Metaformer is actually what you need for vision

    Weihao Yu, Mi Luo, Pan Zhou, Chenyang Si, Yichen Zhou, Xinchao Wang, Jiashi Feng, and Shuicheng Yan. Metaformer is actually what you need for vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10819–10829, 2022. 2

    2022

  42. [50]

    Hivt: Hierarchical vector transformer for multi-agent motion prediction

    Zikang Zhou, Luyao Ye, Jianping Wang, Kui Wu, and Ke- jie Lu. Hivt: Hierarchical vector transformer for multi-agent motion prediction. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 8823–8833, 2022. 1 Efficient Multi-Person Motion...

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.