Pith. sign in

REVIEW 3 major objections 6 minor 46 references

A Structure-aware and Motion-adaptive Framework for 3D Human Pose Estimation with Mamba

T0 review · 3 major / 6 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 SAMA's structure-aware, motion-adaptive Mamba reports 36.5 mm MPJPE on Human3.6M, surpassing PoseMamba-X by 0.6 mm at half the compute.

desk verdict A credible incremental Mamba-based pose-lifting paper whose SOTA claim rests on single-run numbers; the architecture is sound but the evidence needs code, seeds, and variance. read the letter →

arxiv 2507.19852 v1 pith:BTQ7FUF5 submitted 2025-07-26 cs.CV

Ye Lu , Jie Wang , Jianjun Gao , Rui Gong , Chen Cai , Kim-Hui Yap This is my paper
classification cs.CV
keywords 3DhumanposeestimationliftingstatespacemodelsMambastructure-awareintegratormotion-adaptivemodulatortimescalemodulationHuman3.6M
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

Most Mamba-based 3D pose lifters flatten the skeleton into a 1D sequence, which breaks the graph of connected joints, and then push every joint's motion trajectory through the same state-space transition. This paper argues that both choices are the wrong inductive bias: pose topology should be injected into the state space, and each joint should get a timescale matched to its own motion. To that end it introduces SAMA, with a Structure-aware State Integrator that mixes joint features and hidden states through a learned adjacency matrix, and a Motion-adaptive State Modulator that predicts a per-joint timescale from adjacent-frame features. The payoff, if the experiments hold, is a better accuracy-efficiency tradeoff: on Human3.6M, SAMA-L reports 36.5 mm mean per-joint position error (MPJPE) versus 37.1 mm for PoseMamba-X while using about 17.3M parameters and 234M MACs/frame, and 11.4 mm versus 14.8 mm with ground-truth 2D poses. On MPI-INF-3DHP it reports 14.4 mm MPJPE, and ablation shows both modules contribute (37.4 mm with both against 39.9 mm for the bare stack). The paper also shows the two modules can be prepended to existing lifters, cutting MPJPE by 0.6-1.2 mm.

What carries the argument

Two modules on top of a Mamba2 SSD backbone. The Structure-aware State Integrator builds a normalized learnable adjacency matrix $M = \mathrm{softmax}(D^{-1/2}(M_o + I)D^{-1/2})$ and uses it twice: to fuse joint features before the state transition, $x'_a = x_a + \sum_{k} M_{ak} x_k$, and to fuse hidden states after it, $H_a = h_a + \sum_{k} M_{ak} h_k$, before the observation $y_a = C_a H_a$. This replaces expensive bidirectional or global-local scanning with one matrix mix per block. The Motion-adaptive State Modulator predicts a joint-specific timescale from adjacent-frame features, $\Delta_t = S_\Delta(x_t, x_{t-1})$, implemented as a pointwise convolution with kernel size 2 or a concatenation-plus-linear layer; the learned timescale lets fast-moving limb joints weight the current input more and low-motion trunk joints preserve state continuity. Spatial and temporal position embeddings are dropped, on the grounds that the recurrent state already encodes token order. The two modules alternate across $K$ layers, with Jamba-style attention layers included.

What would settle it

Re-run SAMA-L and PoseMamba-X from the same SHNet 2D keypoints, split, and 351-frame windows with at least five random seeds; if the 36.5 vs 37.1 mm MPJPE gap does not persist, or the 11.4 vs 14.8 mm ground-truth-2D gap collapses, the claimed superiority at lower cost is not established.

Watch

Extended reading notes

Core claim

The central claim is that state-space models can do 2D-to-3D pose lifting without sacrificing body structure or joint-specific motion, and that doing so is what makes them both more accurate and cheaper than scanning-based Mamba baselines. In SAMA, the skeleton is not flattened into an arbitrary scan order: a learnable adjacency matrix, initialized from the skeleton graph and normalized by joint degree, is used to aggregate each joint's feature with its neighbors' features before the state transition and to mix the hidden states after it, so outputs depend on connected joints rather than on scan position. The temporal pass then treats joints differently, learning each joint's discretization timescale from the motion of that joint across adjacent frames; high-motion joints receive larger timescales and low-motion trunk joints smaller ones. The measured consequence is state-of-the-art accuracy at reduced cost: 36.5 mm MPJPE and 31.0 mm P-MPJPE on Human3.6M with SHNet 2D poses, 14.4 mm on MPI-INF-3DHP, and 11.4 mm with ground-truth 2D input, all with fewer parameters and MACs than PoseMamba-X.

Load-bearing premise

The evaluation assumes that single-run differences of 0.6-1.2 mm on Human3.6M reflect real model quality rather than run-to-run or protocol variation, since the numbers come from earlier papers without shared code and the paper reports no variance or seed statistics.

Editorial extensions

If this is right

  • A plain vanilla scan with topology-aware state fusion beats bi-directional scanning (38.2 mm) and global-local scanning (37.9 mm) at 37.4 mm MPJPE with fewer MACs, so structure-aware fusion can substitute for costly scanning strategies.
  • Per-joint timescales learned from adjacent frames are a sufficient and cheap mechanism for modeling heterogeneous joint motion, since pointwise convolution (kernel 2) outperforms the linear alternative and the learned timescales correlate with measured motion intensity.
  • SSI and MSM transfer beyond Mamba: prepending them to MixSTE, MotionBERT, and MotionAGFormer lowers MPJPE by 0.6, 1.2, and 0.9 mm respectively, indicating the structure-and-motion inductive bias is architecture-independent.
  • If the ground-truth-2D results (11.4 vs 14.8 mm) hold, the accuracy gain comes from the lifting architecture itself rather than from the upstream 2D detector, strengthening the case for structure-aware SSM lifting in video.

Reading between the lines

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

  • If the learned adjacency matrix is examined across actions, a testable extension is whether it converges to anatomical edges, action-dependent connections, or generic mixing; the paper initializes it from skeleton connectivity but does not analyze the learned pattern.
  • The removal of spatial and temporal position embeddings is asserted without an ablation; a reader can test whether adding them back changes MPJPE, which would tell whether SSM ordering alone really makes them redundant.
  • Because the modules are cheap and model-agnostic, they may transfer to other SSM-based human-centric tasks such as hand pose estimation or motion generation, where joint graphs and heterogeneous motion also matter; the paper demonstrates only pose lifting.
  • A practical corollary the paper leaves implicit is that at 234M MACs/frame, SAMA-L sits in a range where real-time inference on edge hardware may be feasible, provided the 0.6 mm margin over prior art survives reimplementation.
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 / 6 minor

Summary. The paper proposes SAMA, a 2D-to-3D pose-lifting framework built on Mamba2/SSD, with two novel components: a Structure-aware State Integrator (SSI) that fuses joint features and hidden states through a learnable normalized adjacency matrix, and a Motion-adaptive State Modulator (MSM) that conditions the SSM timescale on per-joint motion cues extracted from adjacent frames. The authors report state-of-the-art MPJPE of 36.5 mm on Human3.6M with SHNet 2D poses and 11.4 mm with ground-truth 2D poses, along with 14.4 mm on MPI-INF-3DHP, at lower parameter and MAC counts than PoseMamba-X. Ablations in Tables 3-5 attribute the gains to SSI and MSM, and Table 6 shows consistent improvements when the modules are prepended to MixSTE, MotionBERT, and MotionAGFormer.

Significance. If the reported results are reproducible, the contribution is a meaningful and efficient improvement over prior Mamba-based pose lifters. The idea of injecting pose topology into the state space rather than relying solely on scanning order, and of modulating the SSM timescale with per-joint motion statistics, is well motivated and the ablations support the qualitative claim that each module helps. The generalization results across multiple existing architectures are a particular strength, as they suggest the modules are not overfit to a single backbone. The main limitations are empirical: the headline SOTA differences (0.6 mm over PoseMamba-X; 3.4 mm under ground-truth input) are single-run point estimates with no reported variance, and the paper neither releases code nor specifies all architecture hyperparameters. These gaps prevent verification of the central claim at its current precision.

major comments (3)
  1. [§4.1, Tables 1 and 3] All reported MPJPE numbers are single-run point estimates with no seeds, standard deviations, confidence intervals, or released code/weights. The headline gains of 0.6 mm over PoseMamba-X and 3.4 mm under GT-2D input, as well as the ablation differences 39.9 → 39.3 → 38.4 → 37.4 in Table 3, are within the range of run-to-run and protocol variation typical for Human3.6M. Without repeated runs or statistical tests, the claim that SAMA-L is state of the art at lower cost is not established. Please report multi-seed mean ± std (or equivalent) or release code and train/evaluation scripts so the numbers can be independently verified.
  2. [§3.5, §4.1] The architecture sizes SAMA-S/B/L are never defined: the number of alternating layers K and the feature dimension d are not reported, and the implementation details list only sequence length, optimizer, and learning rate. As a result, the parameter and MAC counts in Table 1 and the efficiency comparison cannot be reproduced or audited. Please specify K, d, and any width/depth differences among the three variants.
  3. [§3.2] The removal of spatial and temporal position embeddings is asserted without an ablation. This is a deliberate departure from prior pose-lifting models and changes the inductive bias of the network; an experiment with and without position embeddings is needed to ensure that the reported gains are attributable to SSI/MSM rather than to this design choice.
minor comments (6)
  1. [§4.2] The text refers to 'PoseMamba-L', but Table 1 lists only PoseMamba-S/B/X; please correct this reference or clarify the intended variant.
  2. [Table 1] The 'vs. prev. SoTA' row is not explained: the parameter reduction should be 9.2M (26.5 → 17.3), not 11.2, and the triple of arrows is ambiguous.
  3. [§4.1] The evaluation metric is described as 'P-MPJPE (%, ↓)' in the text, but Table 1 labels it 'P2: P-MPJPE (mm)'; please make the unit consistent.
  4. [Eq. (4)] D should be defined as a diagonal degree matrix so that D^{-1/2} is meaningful; the current notation treats it as a scalar.
  5. [References] References [12] and [13] are the same PoseMamba paper listed twice; please consolidate them.
  6. [Figure 5] The timescale values in the top row are reported to two decimal places and appear nearly constant across joints; please describe the aggregation (e.g., over frames and actions) and whether these values are averaged.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline MPJPE gains are external test-set measurements and the module ablations are internal comparisons, not fitted constants.

full rationale

Walking the paper's claimed derivation chain, no load-bearing step reduces to its own input by construction. The core modules are defined from input features and learnable parameters: SSI (Eq. 4-6) constructs a learnable adjacency matrix normalized by joint degrees and uses it to fuse joint features and hidden states, while MSM (Eq. 7) learns a joint-specific timescale from adjacent-frame features. Neither module is defined in terms of the target MPJPE or the reported test-set outputs. The headline results in Table 1 (36.5 mm vs. PoseMamba-X 37.1 mm; 11.4 mm vs. 14.8 mm with ground-truth 2D) are measured against the external Human3.6M test protocol, not recovered from a fitted parameter. The ablation in Table 3 compares against an internal vanilla-SSD baseline (39.9 and 39.3 mm), which is a legitimate internal comparison rather than a circular prediction. Figure 5 reads a correlation between limb motion intensity and learned timescales out of the trained model; this is post hoc interpretation supporting the MSM design rationale, not a prediction that is equivalent to its own input. The removal of position embeddings in Section 3.2 is an unablated design choice, and the absence of seeds/variance is a reproducibility weakness, but neither constitutes circularity under the specified criteria. The paper's self-citations (references [1], [20], [21]) appear only in related-work context and are not used as load-bearing evidence for the central claim. Therefore the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities, forces, or conserved quantities are introduced. The learnable adjacency matrix and motion-adaptive timescale are learned network components. The main extras the paper depends on are hand-set hyperparameters, an unstated network depth and width, and several domain assumptions about skeleton topology and motion-timescale correspondence.

free parameters (3)
  • Loss weights lambda_m and lambda_n = lambda_m = 20, lambda_n = 0.5
    Set in Eq. (8) following MotionBERT; no sensitivity analysis is provided, yet these weights directly influence the reported MPJPE values.
  • Number of alternating layers K and feature dimension d = not reported
    The architecture in Section 3.5 uses K alternating SSI/MSM layers plus attention, but K and d are not stated, so parameter counts and MACs cannot be reproduced from the text.
  • Timescale modulator kernel size and padding = kernel size 2, zero padding at the left
    Section 3.4 chooses point-wise convolution over the linear variant with no kernel-size sweep; the choice affects the MSM behavior underlying the main result.
assumptions (4)
  • standard math Zero-Order Hold discretization and Mamba2 State Space Duality are valid background tools
    Section 3.1 applies ZOH discretization and the SSD reformulation from Mamba/Mamba2 without reproving them; these are standard results but unverified inside this paper.
  • domain assumption A fixed human skeleton graph with degree normalization is a useful prior for joint relationships
    Section 3.3, Eq. (4), builds the learnable adjacency matrix from skeleton connectivity and degree normalization; this is the standard GCN prior, not derived in this paper.
  • domain assumption High-motion limb joints should use larger timescales and low-motion trunk joints smaller timescales
    Section 3.4 and Figure 5 motivate MSM with this intensity-timescale correlation. The figure is produced from a trained model, so the assumption is supported post hoc rather than tested before training.
  • domain assumption SSMs already encode token order, so positional embeddings are unnecessary
    Section 3.2 removes spatial and temporal position embeddings based on this assertion. No ablation, citation, or experiment in the paper isolates this design decision.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Structure-aware and Motion-adaptive Framework for 3D Human Pose Estimation with Mamba." pith.science (2026). https://pith.science/paper/BTQ7FUF5

@misc{pith2026250719852,
  author       = {Pith},
  title        = {Pith review of: A Structure-aware and Motion-adaptive Framework for 3D Human Pose Estimation with Mamba},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BTQ7FUF5}},
  note         = {Machine review of arXiv:2507.19852}
}
read the original abstract

Recent Mamba-based methods for the pose-lifting task tend to model joint dependencies by 2D-to-1D mapping with diverse scanning strategies. Though effective, they struggle to model intricate joint connections and uniformly process all joint motion trajectories while neglecting the intrinsic differences across motion characteristics. In this work, we propose a structure-aware and motion-adaptive framework to capture spatial joint topology along with diverse motion dynamics independently, named as SAMA. Specifically, SAMA consists of a Structure-aware State Integrator (SSI) and a Motion-adaptive State Modulator (MSM). The Structure-aware State Integrator is tasked with leveraging dynamic joint relationships to fuse information at both the joint feature and state levels in the state space, based on pose topology rather than sequential state transitions. The Motion-adaptive State Modulator is responsible for joint-specific motion characteristics recognition, thus applying tailored adjustments to diverse motion patterns across different joints. Through the above key modules, our algorithm enables structure-aware and motion-adaptive pose lifting. Extensive experiments across multiple benchmarks demonstrate that our algorithm achieves advanced results with fewer computational costs.

Figures

Figures reproduced from arXiv: 2507.19852 by the authors.

Figure 1
Figure 1. (a) Illustration of structure-aware state integrator. On top [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Comparisons of various 3D Human Pose Estimation [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The overview of our proposed SAMA. (a): Our Network Structure. The core part is the alternative stack of structure-aware state [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Statistical motion intensity and timescale ∆ results across different joints. 0.04 0.04 0.04 0.04 0.04 0.04 0.04 0.04 0.03 0.03 0.03 0.03 0.03 0.02 0.02 0.01 0.02 0.17 0.14 0.14 0.18 0.12 0.14 0.15 0.11 0.13 0.10 0.09 0.09 0.10 0.08 0.09 0.06 0.05 joints on limbs joint…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [13]

    PoseMamba: Monocular 3D Human Pose Estimation with Bidirectional Global-Local Spatio-Temporal State Space Model

    Yunlong Huang, Junshuo Liu, Ke Xian, and Robert Caim- ing Qiu. Posemamba: Monocular 3d human pose estimation with bidirectional global-local spatio-temporal state space model. arXiv preprint arXiv:2408.03540, 2024

    work page Pith review arXiv 2024

  2. [1]

    Empowering Large Language Model for Continual Video Question Answering with Collaborative Prompting

    Chen Cai, Zheng Wang, Jianjun Gao, Wenyang Liu, Ye Lu, Runzhong Zhang, and Kim-Hui Yap. Empowering large lan- guage model for continual video question answering with collaborative prompting. arXiv preprint arXiv:2410.00771, 2024

    work page Pith review arXiv

  3. [2]

    Hdformer: High-order directed trans- former for 3d human pose estimation

    Hanyuan Chen, Jun-Yan He, Wangmeng Xiang, Zhi-Qi Cheng, Wei Liu, Hanbing Liu, Bin Luo, Yifeng Geng, and Xuansong Xie. Hdformer: High-order directed trans- former for 3d human pose estimation. arXiv preprint arXiv:2302.01825, 2023

    arXiv 2023

  4. [3]

    Anatomy-aware 3d human pose estimation with bone-based pose decomposition

    Tianlang Chen, Chen Fang, Xiaohui Shen, Yiheng Zhu, Zhili Chen, and Jiebo Luo. Anatomy-aware 3d human pose estimation with bone-based pose decomposition. IEEE Transactions on Circuits and Systems for Video Technology, 32(1):198–209, 2021

    work page 2021

  5. [4]

    Deductive learn- ing for weakly-supervised 3d human pose estimation via un- calibrated cameras

    Xipeng Chen, Pengxu Wei, and Liang Lin. Deductive learn- ing for weakly-supervised 3d human pose estimation via un- calibrated cameras. In Thirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial...

    work page 2021

  6. [5]

    So-handnet: Self-organizing network for 3d hand pose estimation with semi-supervised learning

    Yujin Chen, Zhigang Tu, Liuhao Ge, Dejun Zhang, Ruizhi Chen, and Junsong Yuan. So-handnet: Self-organizing network for 3d hand pose estimation with semi-supervised learning. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6960–6969. IEEE, 2019

    work page 2019

  7. [6]

    Joint hand-object 3d reconstruction from a single image with cross-branch fea- ture fusion

    Yujin Chen, Zhigang Tu, Di Kang, Ruizhi Chen, Linchao Bao, Zhengyou Zhang, and Junsong Yuan. Joint hand-object 3d reconstruction from a single image with cross-branch fea- ture fusion. IEEE Trans. Image Process., 30:4008–4021, 2021

    work page 2021

  8. [7]

    Transformers are ssms: General- ized models and efficient algorithms through structured state space duality

    Tri Dao and Albert Gu. Transformers are ssms: General- ized models and efficient algorithms through structured state space duality. In Forty-first International Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27,

    work page 2024

Show all 46 references
  1. [8]

    Hamba: Single- view 3d hand reconstruction with graph-guided bi-scanning mamba

    Haoye Dong, Aviral Chharia, Wenbo Gou, Francisco Vi- cente Carrasco, and Fernando De la Torre. Hamba: Single- view 3d hand reconstruction with graph-guided bi-scanning mamba. In Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information Proce...

    2024

  2. [9]

    Uplift and upsample: Efficient 3d human pose estimation with up- lifting transformers

    Moritz Einfalt, Katja Ludwig, and Rainer Lienhart. Uplift and upsample: Efficient 3d human pose estimation with up- lifting transformers. In IEEE/CVF Winter Conference on Applications of Computer Vision, W ACV2023, Waikoloa, HI, USA, January 2-7, 2023, pages 2902–2912. IEEE, 2023

    2023

  3. [10]

    Mamba: Linear-time sequence mod- eling with selective state spaces

    Albert Gu and Tri Dao. Mamba: Linear-time sequence mod- eling with selective state spaces. CoRR, abs/2312.00752, 2023

    2023 arXiv

  4. [11]

    Combining recurrent, convolutional, and continuous-time models with linear state space layers

    Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher R ´e. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572–585, 2021

    2021

  5. [14]

    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

    2013

  6. [15]

    Pose-oriented transformer with uncertainty-guided refinement for 2d-to-3d human pose esti- mation

    Han Li, Bowen Shi, Wenrui Dai, Hongwei Zheng, Bo- tao Wang, Yu Sun, Min Guo, Chenglin Li, Junni Zou, and Hongkai Xiong. Pose-oriented transformer with uncertainty-guided refinement for 2d-to-3d human pose esti- mation. In Proceedings of the AAAI conference on artificial intell...

    2023

  7. [16]

    Sijin Li, Weichen Zhang, and Antoni B. Chan. Maximum- margin structured learning with deep networks for 3d human pose estimation. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 2848–2856. IEEE Computer Society, 2015

    2015

  8. [17]

    Mhformer: Multi-hypothesis transformer for 3d human pose estimation

    Wenhao Li, Hong Liu, Hao Tang, Pichao Wang, and Luc Van Gool. Mhformer: Multi-hypothesis transformer for 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 13137–13146. IEEE, 2022

    2022

  9. [18]

    Exploiting temporal contexts with strided transformer for 3d human pose estimation.IEEE Trans

    Wenhao Li, Hong Liu, Runwei Ding, Mengyuan Liu, Pichao Wang, and Wenming Yang. Exploiting temporal contexts with strided transformer for 3d human pose estimation.IEEE Trans. Multim., 25:1282–1293, 2023

    2023

  10. [19]

    Jamba: A hybrid transformer-mamba language model

    Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, et al. Jamba: A hybrid transformer-mamba language model. arXiv preprint arXiv:2403.19887, 2024

    2024 arXiv

  11. [20]

    Bitstream-corrupted video recov- ery: A novel benchmark dataset and method

    Tianyi Liu, Kejun Wu, Yi Wang, Wenyang Liu, Kim-Hui Yap, and Lap-Pui Chau. Bitstream-corrupted video recov- ery: A novel benchmark dataset and method. Advances in Neural Information Processing Systems, 2023

    2023

  12. [21]

    Hdplifter: Hierarchical dynamics per- ception for 2d-to-3d human pose lifting

    Ye Lu, Jianjun Gao, Chen Cai, Ruoyu Wang, Duc Tri Phan, and Kim-Hui Yap. Hdplifter: Hierarchical dynamics per- ception for 2d-to-3d human pose lifting. In 2024 IEEE International Conference on Image Processing (ICIP). IEEE, 2024

    2024

  13. [22]

    Mo- tionagformer: Enhancing 3d human pose estimation with a transformer-gcnformer network

    Soroush Mehraban, Vida Adeli, and Babak Taati. Mo- tionagformer: Enhancing 3d human pose estimation with a transformer-gcnformer network. In IEEE/CVF Winter 9 Conference on Applications of Computer Vision, W ACV 2024, Waikoloa, HI, USA, January 3-8, 2024, pages 6905–

    2024

  14. [23]

    Monocular 3d human pose estimation in the wild using improved cnn supervision

    Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3d human pose estimation in the wild using improved cnn supervision. In 2017 international conference on 3D vision (3DV), pages 506–516. IEEE, 2017

    2017

  15. [24]

    I2l-meshnet: Image- to-lixel prediction network for accurate 3d human pose and mesh estimation from a single RGB image

    Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image- to-lixel prediction network for accurate 3d human pose and mesh estimation from a single RGB image. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VII, pages 752–

    2020

  16. [25]

    Derpa- nis, and Kostas Daniilidis

    Georgios Pavlakos, Xiaowei Zhou, Konstantinos G. Derpa- nis, and Kostas Daniilidis. Coarse-to-fine volumetric pre- diction for single-image 3d human pose. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages ...

    2017

  17. [26]

    Jihua Peng, Yanghong Zhou, and P. Y . Mok. Ktpformer: Kinematics and trajectory prior knowledge-enhanced trans- former for 3d human pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, W A,USA, June 16-22, 2024, pages 1123–113...

    2024

  18. [27]

    P-STMO: pre-trained spatial temporal many-to-one model for 3d human pose estimation

    Wenkang Shan, Zhenhua Liu, Xinfeng Zhang, Shanshe Wang, Siwei Ma, and Wen Gao. P-STMO: pre-trained spatial temporal many-to-one model for 3d human pose estimation. In Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv,Israel, October 23-27, 2022, Proceedings, Par...

    2022

  19. [28]

    Integral human pose regression

    Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VI, pages 536–553. Springer, 2018

    2018

  20. [29]

    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 IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 4790–

    2023

  21. [30]

    Direct prediction of 3d body poses from motion com- pensated sequences

    Bugra Tekin, Artem Rozantsev, Vincent Lepetit, and Pascal Fua. Direct prediction of 3d body poses from motion com- pensated sequences. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas,NV , USA, June 27-30, 2016, pages 991–1000. IEEE Comp...

    2016

  22. [31]

    Probabilistic monocular 3d human pose estimation with normalizing flows

    Tom Wehrbein, Marco Rudolph, Bodo Rosenhahn, and Bastian Wandt. Probabilistic monocular 3d human pose estimation with normalizing flows. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages 11179– 11188. IEEE, 2021

    2021

  23. [32]

    Graph stacked hourglass networks for 3d human pose estimation

    Tianhan Xu and Wataru Takano. Graph stacked hourglass networks for 3d human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16105–16114, 2021

    2021

  24. [33]

    Mang Ye, He Li, Bo Du, Jianbing Shen, Ling Shao, and Steven C. H. Hoi. Collaborative refining for person re- identification with label noise. IEEE Trans. Image Process., 31:379–391, 2022

    2022

  25. [34]

    Bruce X. B. Yu, Zhi Zhang, Yongxu Liu, Sheng-Hua Zhong, Yan Liu, and Chang Wen Chen. GLA-GCN: global-local adaptive graph convolutional network for 3d human pose es- timation from monocular video. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, Franc...

    2023

  26. [35]

    Unsupervised pre-training for tem- poral action localization tasks

    Can Zhang, Tianyu Yang, Junwu Weng, Meng Cao, Jue Wang, and Yuexian Zou. Unsupervised pre-training for tem- poral action localization tasks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 14011–14021....

    2022

  27. [36]

    Mixste: Seq2seq mixed spatio-temporal en- coder for 3d human pose estimation in video

    Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Junsong Yuan. Mixste: Seq2seq mixed spatio-temporal en- coder for 3d human pose estimation in video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, page...

    2022

  28. [37]

    A spatial attentive and temporal dilated (SATD) GCN for skeleton-based action recognition

    Jiaxu Zhang, Gaoxiang Ye, Zhigang Tu, Yongtao Qin, Qian- qing Qin, Jinlu Zhang, and Jun Liu. A spatial attentive and temporal dilated (SATD) GCN for skeleton-based action recognition. CAAI Trans. Intell. Technol., 7(1):46–55, 2022

    2022

  29. [38]

    Pose magic: Efficient and temporally consis- tent human pose estimation with a hybrid mamba-gcn net- work

    Xinyi Zhang, Qiqi Bao, Qinpeng Cui, Wenming Yang, and Qingmin Liao. Pose magic: Efficient and temporally consis- tent human pose estimation with a hybrid mamba-gcn net- work. CoRR, abs/2408.02922, 2024

    2024 arXiv

  30. [39]

    Reid, Richard I

    Zeyu Zhang, Akide Liu, Ian D. Reid, Richard I. Hartley, Bohan Zhuang, and Hao Tang. Motion mamba: Efficient and long sequence motion generation. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part I, pages 26...

    2024

  31. [40]

    Long Zhao, Xi Peng, Yu Tian, Mubbasir Kapadia, and Dim- itris N. Metaxas. Semantic graph convolutional networks for 3d human pose regression. InIEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 3425–3435. Comput...

    2019

  32. [41]

    Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation

    Qitao Zhao, Ce Zheng, Mengyuan Liu, Pichao Wang, and Chen Chen. Poseformerv2: Exploring frequency domain for efficient and robust 3d human pose estimation. InIEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pag...

    2023

  33. [42]

    Graformer: Graph-oriented transformer for 3d pose estimation

    Weixi Zhao, Weiqiang Wang, and Yunjie Tian. Graformer: Graph-oriented transformer for 3d pose estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 20406–20415. IEEE, 2022

    2022

  34. [43]

    3d human pose estimation 10 with spatial and temporal transformers

    Ce Zheng, Sijie Zhu, Mat ´ıas Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation 10 with spatial and temporal transformers. In 2021 IEEE/CVF International Conference on Computer Vision, ICCV 2021, Montreal, QC, Canada, October 10-17, 2021, pages...

    2021

  35. [44]

    3d human pose estimation with spatial and temporal transformers

    Ce Zheng, Sijie Zhu, Matias Mendieta, Taojiannan Yang, Chen Chen, and Zhengming Ding. 3d human pose estimation with spatial and temporal transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11656–11665, 2021

    2021

  36. [45]

    Hipart: Hier- archical pose autoregressive transformer for occluded 3d hu- man pose estimation

    Hongwei Zheng, Han Li, Wenrui Dai, Ziyang Zheng, Chenglin Li, Junni Zou, and Hongkai Xiong. Hipart: Hier- archical pose autoregressive transformer for occluded 3d hu- man pose estimation. In Proceedings of the Computer Vision and Pattern Recognition Conference, 2025

    2025

  37. [46]

    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 IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 15039–15053. IEE...

    2023

  38. [2024]

    OpenReview.net, 2024

    2024

Pith tools

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