Pith. sign in

REVIEW 3 major objections 7 minor 58 references

Learning Pyramid-structured Long-range Dependencies for 3D Human Pose Estimation

T0 review · 3 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Pyramid attention drops 3D pose error to 34.2 mm

desk verdict A simple, plausible pyramid-attention twist on GraFormer with consistent but small MPJPE gains; the main claim likely holds, but the thin margins and several printed inconsistencies need refereeing. read the letter →

arxiv 2506.02853 v1 pith:OTENWNCP submitted 2025-06-03 cs.CV

Mingjie Wei , Xuemei Xie , Yutong Zhong , Guangming Shi This is my paper
classification cs.CV
keywords 3Dhumanposeestimationlong-rangedependencepyramidstructureself-attentiongraphconvolutionalnetworktransformermulti-scalefusionbodyhierarchy
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 argues that the long-range dependencies which help reconstruct a 3D pose from 2D joints are best modeled not between individual joints alone, but across a hierarchy of body parts and regions. To capture this, it proposes Pyramid Graph Attention (PGA), which pools joints into part-level and region-level features, appends them to the original joint sequence, and feeds the merged sequence into the keys and values of multi-head self-attention. The resulting network, PGFormer, reports 34.2 mm MPJPE on Human3.6M with ground-truth 2D input and 49.5 mm with detected 2D joints, improving on the 35.2 mm and 51.8 mm of its GraFormer baseline while using roughly the same parameter count. The same PGA design is also inserted into a diffusion-based estimator, DiffPyramid, which reports 31.0 mm. If these results hold, they show that hierarchical structure can be injected into attention cheaply, without deepening the network or multiplying branches.

What carries the argument

The central mechanism is the Pyramid Graph Attention (PGA) module: a multi-head self-attention layer whose keys and values are computed from a pyramid of pooled joint features rather than from the raw joint sequence alone. Two average-pooling operations, P1 and P2, are defined over hand-chosen anatomical groups, reducing 16 joints to 8 part-level nodes and 4 region-level nodes; after layer normalization the pooled features are concatenated with the original joints, and this concatenated sequence is projected to K and V, while the query Q stays on the original joints. This design makes the attention matrix of size (N+M1+M2) x N, adding only a small overhead while letting each joint attend to part and region summaries. In the full PGFormer, the PGA module is paired with a Chebyshev graph convolution layer that replaces the MLP of a standard transformer block, and residual connections wrap both. A diffusion variant, DiffPyramid, reuses the same PGFormer as its denoising network, conditioning the reverse process on 2D pose features.

What would settle it

Replace the anatomical pooling groups in P1 and P2 with randomly permuted groupings of the same sizes, retrain on Human3.6M, and compare MPJPE; if a random grouping matches the 34.5 mm of the pyramid ablation, the human-structure prior is not the cause of the gain.

Watch

Extended reading notes

Core claim

The central claim is that capturing correlations between joints and their pooled body-part and body-region summaries is sufficient to model the long-range dependencies needed for accurate 3D pose estimation from a single frame. Instead of learning one global attention over all joints (which introduces uncorrelated noise) or hand-crafting specific joint-pair connections, the Pyramid Graph Attention module constructs a pyramid of pooled features: P1 merges 16 (or 17) joints into 8 part-level nodes and P2 merges them into 4 region-level nodes. These pooled features, after layer normalization and concatenation with the original joint features, serve as the keys and values of the attention, while the queries remain the original joints. The attention then computes, in parallel, correlations across scales, letting each joint be constrained by the body parts and regions it belongs to. The paper supports this with ablations showing that the pyramid structure lowers MPJPE from 35.2 mm to 34.5 mm on ground-truth input, that manual anatomical pooling beats adaptive graph pooling, and that the PGA module improves several graph-convolution backbones when used as a plug-in.

Load-bearing premise

The performance gain rests on the assumption that the hand-chosen anatomical grouping of joints into eight part-level and four region-level nodes is the right hierarchy for every action and every skeleton.

Editorial extensions

If this is right

  • Single-frame 3D pose estimation can reach a 34.2 mm average MPJPE on Human3.6M with ground-truth 2D joints, improving on the 35.2 mm of GraFormer and the 35.8 mm of GraphSH while using 0.66M parameters.
  • With CPN-detected 2D joints, PGFormer reports 49.5 mm MPJPE, a relative reduction of about 4% over the 51.8 mm GraFormer baseline, with no pose refinement or temporal information.
  • The PGA module is plug-and-play: replacing the long-range component of SemGCN, Modulated GCN, or ChebGCN with PGA reduces MPJPE for each backbone (e.g., from 52.5 mm to 38.2 mm for SemGCN).
  • DiffPyramid, which inserts the same hierarchical spatial constraint into a diffusion model, reaches 31.0 mm MPJPE, outperforming the single-frame diffusion baseline DiffPose (31.6 mm).
  • The architecture runs at about 3200 frames per second on a batch of 64 on the reported GPU, roughly 10 times faster than the diffusion-based method, supporting real-time application.

Reading between the lines

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

  • The fixed anatomical pooling groups are a strong prior, so if a different skeleton definition invalidates those groups, the attention keys and values are polluted before learning begins and the reported gains may not transfer.
  • The efficiency of using a few pooled summaries as keys and values points to a general recipe for injecting global context into attention models for structured prediction; testing it on hand or animal pose could reveal whether the mechanism is specific to the human body.
  • The ablation showing manual grouping beats adaptive pooling suggests the anatomical prior, not merely extra parameters, drives the improvement; a natural extension is to learn the pooling groups from data while initializing from anatomical groups.
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 / 7 minor

Summary. The paper proposes a pyramid-structured graph attention (PGA) module that pools human joint features into part- and region-level nodes and uses these hierarchical features as keys and values in multi-head self-attention, alongside graph convolution layers, forming PGFormer for single-frame 3D human pose estimation. It also introduces DiffPyramid, a diffusion-based variant that uses PGFormer as a conditional spatial prior. The claimed contributions are improved MPJPE on Human3.6M (34.2 mm with ground-truth 2D input, 49.5 mm with CPN detections) compared with GraFormer (35.2 mm and 51.8 mm), better cross-dataset generalization on MPI-INF-3DHP, and a favorable accuracy/parameter trade-off. The paper includes ablations on pooling design, feature fusion, graph convolution backbones, and a plug-and-play PGA module.

Significance. If the empirical results hold, the paper makes a useful contribution to efficient 3D HPE by showing that a lightweight pyramid pooling scheme can encode human sub-structure in self-attention without large parameter growth. The authors provide a public code release, a complexity analysis, and plug-and-play experiments across GCN and diffusion architectures, which are strengths. However, the headline improvements over the direct baselines are small (1.0 mm and 0.6 mm), and the paper's own repeated-run result shows a 0.4 mm spread; without multi-seed statistics the central empirical claim is not yet secure.

major comments (3)
  1. [Section III-B, Eq. (2)] The Chebyshev polynomial recurrence is stated with T0=0, T1=x. Standard Chebyshev polynomials of the first kind require T0(x)=1; with T0=0 the recurrence gives T2(x)=2x^2 instead of 2x^2-1, and the k=0 term in the sum is zero, so Eq. (2) does not implement a Chebyshev filter as defined in [45]. Please correct the definition and clarify which recurrence the released code actually uses.
  2. [Section IV-D, Tables I-II and VII] The main comparisons rest on small margins (34.2 vs 35.2 mm on GT; 31.0 vs 31.6 mm for DiffPyramid), yet the only repeated-run evidence reported, Table VII, shows a 0.4 mm spread for a single pooling configuration. No standard deviations or multiple-seed results are given for the main tables. Please provide seed-averaged errors (mean±std) for at least the primary comparisons and state whether the improvements are statistically significant.
  3. [Table X vs Tables II and XI] The SemGCN baseline in Table X is 52.5 mm with 0.27M parameters, while Table XI reports SemGCN SeqRes at 40.7 mm (0.43M) and Table II reports SemGCN at 43.8 mm. The text says the baseline was run without the non-local module and with hidden dimensions adjusted to match GraphSH, but this does not explain a gap of 9–12 mm. The reported 27% relative improvement from adding PGA appears inflated by the weak baseline. Please report a consistent SemGCN configuration or justify the chosen setting explicitly.
minor comments (7)
  1. [Eq. (1)] The matrix \tilde{A} is defined as I - D^{-1/2}AD^{-1/2}, which is the normalized graph Laplacian, not a normalized adjacency matrix; please adjust the terminology to avoid confusion.
  2. [Section IV-D, Table VII] The text states that for scale 4, the repeated experiment ranged between 35.4 and 35.0, but Table VII lists a single value (35.0); specify which grouping configuration is reported and list the range in the table.
  3. [Throughout] There are several typos: 'Beacause' in Section IV-D, 'effciently' in the Introduction, 'formoulated' in Section III-D, 'formulaated' in Section III-D, and 'feets' in Section IV-E.
  4. [Section IV-C, Table I] The claim that PGFormer is 'significant' relative to GraFormer is stated without statistical support; rephrase to avoid implying significance without evidence.
  5. [Supplementary Material, Eq. (1)] The complexity notation uses dk, N, and Mi without a clear definition of which dimensions correspond to which figures; please define these terms in the main text or supplement.
  6. [Fig. 2] The caption states parameter reductions of 83% and 78% relative to GraphSH and PHGANet, but these are relative reductions in model size, not error; the caption should clarify that the comparison is about parameters, not MPJPE.
  7. [Table XII] The row 'GraFormer [8] - 35.6' is ambiguous; specify which spatial condition is used for the GraFormer baseline.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports measured benchmark errors, not derived predictions, and no load-bearing self-citation chain is present.

full rationale

The paper's central claims are empirical: PGA-based PGFormer and DiffPyramid achieve certain MPJPE values on Human3.6M and MPI-INF-3DHP. These numbers are obtained by training on the stated train subjects and evaluating on the held-out test subjects under Protocol #1, compared against externally published results (e.g., GraFormer at 35.2/51.8 mm, DiffPose at 31.6 mm). No test error is obtained by plugging the model's own output back into the definition of the target; the loss functions (Eqs. 17 and 18) minimize distance to ground-truth 3D poses, and the reported quantities are independent measurements on the test set, not fitted parameters relabeled as predictions. The PGA module is defined by pooling operations (Eq. 3), concatenation (Eq. 4), and attention over key/value features derived from the pyramid (Eqs. 6-7); these definitions do not presuppose the benchmark outcomes. The one co-authored citation (ref. [26] in the related-work discussion of temporal methods) is not load-bearing. The pooling grouping is hand-designed and could be questioned on transfer grounds, and the small margins relative to the reported 0.4 mm run-to-run spread (Section IV-D) are a legitimate statistical-robustness concern, but variability in empirical evaluation is not a circularity. There is no self-citation invoked as an external uniqueness theorem, no ansatz smuggled in via the authors' own prior work, and no renaming of a known result as unification. The paper is self-contained in the relevant sense: its conclusions rest on benchmark measurements and ablations, not on definitions or self-referential support.

Assumptions & free parameters 6 free parameters · 3 assumptions · 0 invented entities

The central claim rests on standard benchmark evaluation plus a hand-designed skeleton hierarchy; it does not rest on any formal derivation or new physical assumption. The main non-empirical inputs are hyperparameters and the manual pooling groups, none of which are derived from first principles.

free parameters (6)
  • Lambda (L1 regularization weight) = 0.1 (CPN input), 0.025 (ground-truth input)
    Eq. (17) mixes L1 and L2 pose losses; lambda is selected by the ablation in Table XIII rather than derived.
  • Hidden layer size = 256 (CPN), 96 (ground-truth)
    Set in Section IV-B for convergence; no sensitivity analysis is reported.
  • Manual pooling groups P1/P2 = P1: 16 or 17 joints to 8 parts; P2: to 4 regions
    Eq. (3) and Fig. 7 define the groups by hand; Table VII shows they outperform adaptive pooling methods.
  • Number of attention heads = 4
    Supplementary implementation details state 4 heads; the value is not varied in the paper.
  • Dropout rates = 0.25 after GCN/attention, 0.05 after softmax
    Supplementary implementation details; regularization hyperparameters chosen by hand.
  • Diffusion noise schedule and sampling = beta from 1e-4 to 2e-3, reverse steps K=50, samples N=5
    Section IV-B; standard DDPM/DDIM settings used without derivation.
assumptions (3)
  • domain assumption 2D-to-3D lifting is a valid decomposition of the pose problem
    The method takes ground-truth or CPN 2D keypoints as input and regresses 3D coordinates, following [7]-[11]; the paper does not justify this reduction.
  • ad hoc to paper Manual pooling groups capture human part and region semantics
    Eq. (3)-(4) and Fig. 7 define P1 and P2 by hand; if the groups do not correspond to real coordination patterns, the cross-scale keys and values mix irrelevant joints.
  • domain assumption Benchmark protocols measure pose quality
    MPJPE on Human3.6M and 3DPCK/AUC on MPI-INF-3DHP are accepted without argument, and comparisons rely on numbers reported in prior papers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Pyramid-structured Long-range Dependencies for 3D Human Pose Estimation." pith.science (2026). https://pith.science/paper/OTENWNCP

@misc{pith2026250602853,
  author       = {Pith},
  title        = {Pith review of: Learning Pyramid-structured Long-range Dependencies for 3D Human Pose Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OTENWNCP}},
  note         = {Machine review of arXiv:2506.02853}
}
read the original abstract

Action coordination in human structure is indispensable for the spatial constraints of 2D joints to recover 3D pose. Usually, action coordination is represented as a long-range dependence among body parts. However, there are two main challenges in modeling long-range dependencies. First, joints should not only be constrained by other individual joints but also be modulated by the body parts. Second, existing methods make networks deeper to learn dependencies between non-linked parts. They introduce uncorrelated noise and increase the model size. In this paper, we utilize a pyramid structure to better learn potential long-range dependencies. It can capture the correlation across joints and groups, which complements the context of the human sub-structure. In an effective cross-scale way, it captures the pyramid-structured long-range dependence. Specifically, we propose a novel Pyramid Graph Attention (PGA) module to capture long-range cross-scale dependencies. It concatenates information from various scales into a compact sequence, and then computes the correlation between scales in parallel. Combining PGA with graph convolution modules, we develop a Pyramid Graph Transformer (PGFormer) for 3D human pose estimation, which is a lightweight multi-scale transformer architecture. It encapsulates human sub-structures into self-attention by pooling. Extensive experiments show that our approach achieves lower error and smaller model size than state-of-the-art methods on Human3.6M and MPI-INF-3DHP datasets. The code is available at https://github.com/MingjieWe/PGFormer.

Figures

Figures reproduced from arXiv: 2506.02853 by the authors.

Figure 1
Figure 1. Top: Illustration of long-range dependence differences among global aggregation [7], [14], hand-crafted connections [15] and ours. The proposed method learns potential correlation with fewer unrelated noise. Joints should not only be constrained by other individual joints but also be modulated by the body parts. Bottom: We propose to capture long-range dependencies via pyramid structure. The relationship can be lear… view at source ↗
Figure 2
Figure 2. Comparison of the performance and model size between the proposed [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The network architecture of the proposed [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Architecture and illustration of Pyramid Graph Attention. We construct [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Architecture of the proposed DiffPyramid for 3D HPE. On the left, [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: MPJPE distribution on the testset of Human3.6M. We compare [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: A diagram illustrating the pooling strategy and construction of pyramid [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: A diagram displaying the other scales of pooling. [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Qualitative comparisons on Human3.6M dataset. Ground truth is the [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 11
Figure 11. Figure 11: Visualization of attention maps in different actions. The proposed [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Visualization on various in-the-wild scenarios. These challenge scenarios of varying difficulty demonstrate the effectiveness and robustness of our [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 1
Figure 1. Figure 1: Top: Diagram illustrating long-range dependencies learned by a PGA module without l1 − norm. Bottom: Corresponding diagram showing long￾range dependencies learned by a PGA module with l1 − norm. As shown in [PITH_FULL_IMAGE:figures/full_fig_p015_1.png]
Figure 2
Figure 2. Figure 2: A diagram illustrating the pooling of 17 keypoints. The ’Nose’ [PITH_FULL_IMAGE:figures/full_fig_p016_2.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

58 extracted references · 57 canonical work pages

  1. [45]

    Defferrard, X

    M. Defferrard, X. Bresson, and P. Vandergheynst, ``Convolutional neural networks on graphs with fast localized spectral filtering,'' Advances in Neural Information Processing Systems, vol. 29, 2016

    work page 2016

  2. [1]

    C. Pang, X. Lu, and L. Lyu, `` Skeleton-Based Action Recognition Through Contrasting Two-Stream Spatial-Temporal Networks ,'' IEEE Transactions on Multimedia, vol. 25, pp. 8699--8711, 2023

    work page 2023

  3. [2]

    Angelini, Z

    F. Angelini, Z. Fu, Y. Long, L. Shao, and S. M. Naqvi, `` 2D Pose-Based Real-Time Human Action Recognition With Occlusion-Handling ,'' IEEE Transactions on Multimedia, vol. 22, no. 6, pp. 1433--1446, 2020

    work page 2020

  4. [3]

    S. Tu, Q. Dai, Z. Wu, Z.-Q. Cheng, H. Hu, and Y.-G. Jiang, ``Implicit temporal modeling with learnable alignment for video recognition,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 19\,936--19\,947

    work page 2023

  5. [4]

    Q. Bao, W. Liu, Y. Cheng, B. Zhou, and T. Mei, `` Pose-Guided Tracking-by-Detection: Robust Multi-Person Pose Tracking ,'' IEEE Transactions on Multimedia, vol. 23, pp. 161--175, 2021

    work page 2021

  6. [5]

    Mehta, S

    D. Mehta, S. Sridhar, O. Sotnychenko, H. Rhodin, M. Shafiei, H.-P. Seidel, W. Xu, D. Casas, and C. Theobalt, `` VNect: Real-time 3D Human Pose Estimation with a Single RGB Camera ,'' ACM Transactions on Graphics, vol. 36, no. 4, 2017

    work page 2017

  7. [6]

    Y. Zhou, X. Yan, Z.-Q. Cheng, Y. Yan, Q. Dai, and X.-S. Hua, `` BlockGCN: Redefining Topology Awareness for Skeleton-Based Action Recognition ,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024

    work page 2024

  8. [7]

    Zou and W

    Z. Zou and W. Tang, ``Modulated graph convolutional network for 3D human pose estimation,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 11\,477--11\,487

    work page 2021

Show all 58 references
  1. [8]

    W. Zhao, W. Wang, and Y. Tian, `` GraFormer: Graph-oriented Transformer for 3D Pose Estimation ,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 20\,438--20\,447

    2022

  2. [9]

    Y. Cai, L. Ge, J. Liu, J. Cai, T.-J. Cham, J. Yuan, and N. M. Thalmann, ``Exploiting spatial-temporal relationships for 3D pose estimation via graph convolutional networks,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 2272--2281

    2019

  3. [10]

    Martinez, R

    J. Martinez, R. Hossain, J. Romero, and J. J. Little, ``A simple yet effective baseline for 3D human pose estimation,'' in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 2640--2649

    2017

  4. [11]

    L. Zhao, X. Peng, Y. Tian, M. Kapadia, and D. N. Metaxas, `` Semantic Graph Convolutional Networks for 3D Human Pose Regression ,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 3425--3435

    2019

  5. [12]

    J. Gong, L. G. Foo, Z. Fan, Q. Ke, H. Rahmani, and J. Liu, `` DiffPose: Toward More Reliable 3D Pose Estimation ,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 13\,041--13\,051

    2023

  6. [13]

    Chen, J.-Y

    H. Chen, J.-Y. He, W. Xiang, Z.-Q. Cheng, W. Liu, H. Liu, B. Luo, Y. Geng, and X. Xie, `` HDFormer: High-order Directed Transformer for 3D Human Pose Estimation ,'' in Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, 2023, pp. 581--589

    2023

  7. [14]

    Zheng, S

    C. Zheng, S. Zhu, M. Mendieta, T. Yang, C. Chen, and Z. Ding, `` 3D human pose estimation with spatial and temporal transformers,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 11\,656--11\,665

    2021

  8. [15]

    H.-S. Fang, Y. Xu, W. Wang, X. Liu, and S.-C. Zhu, ``Learning pose grammar to encode human body configuration for 3D pose estimation,'' in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, 2018

    2018

  9. [16]

    Xu and W

    T. Xu and W. Takano, `` Graph Stacked Hourglass Networks for 3D Human Pose Estimation ,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 16\,105--16\,114

    2021

  10. [17]

    W. Hu, C. Zhang, F. Zhan, L. Zhang, and T.-T. Wong, ``Conditional directed graph convolution for 3D human pose estimation,'' in Proceedings of the 29th ACM International Conference on Multimedia, 2021, pp. 602--611

    2021

  11. [18]

    H. Ci, C. Wang, X. Ma, and Y. Wang, ``Optimizing network structure for 3D human pose estimation,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 2262--2271

    2019

  12. [19]

    K. Liu, R. Ding, Z. Zou, L. Wang, and W. Tang, ``A comprehensive study of weight sharing in graph networks for 3D human pose estimation,'' in Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part X 16. 1em plus 0.5em minus 0....

    2020

  13. [20]

    A. Zeng, X. Sun, L. Yang, N. Zhao, M. Liu, and Q. Xu, ``Learning skeletal graph neural networks for hard 3D pose estimation,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 11\,436--11\,445

    2021

  14. [21]

    Zhang, C

    S. Zhang, C. Wang, L. Nie, H. Yao, Q. Huang, and Q. Tian, ``Learning enriched hop-aware correlation for robust 3D human pose estimation,'' International Journal of Computer Vision, vol. 131, no. 6, pp. 1566--1583, 2023

    2023

  15. [22]

    Ionescu, D

    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, 2014

    2014

  16. [23]

    Lee and Z

    H.-J. Lee and Z. Chen, ``Determination of 3D human body postures from a single view,'' Computer Vision, Graphics, and Image Processing, vol. 30, no. 2, pp. 148--168, 1985

    1985

  17. [24]

    Mehta, H

    D. Mehta, H. Rhodin, D. Casas, P. Fua, O. Sotnychenko, W. Xu, and C. Theobalt, ``Monocular 3D human pose estimation in the wild using improved cnn supervision,'' in 2017 International Conference on 3D Vision . 1em plus 0.5em minus 0.4em IEEE, 2017, pp. 506--516

    2017

  18. [25]

    Pavlakos, X

    G. Pavlakos, X. Zhou, K. G. Derpanis, and K. Daniilidis, ``Coarse-to-fine volumetric prediction for single-image 3D human pose,'' in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 7025--7034

    2017

  19. [26]

    Zhong, G

    Y. Zhong, G. Yang, D. Zhong, X. Yang, and S. Wang, `` Frame-Padded Multiscale Transformer for Monocular 3D Human Pose Estimation ,'' IEEE Transactions on Multimedia, vol. 26, pp. 6191--6201, 2024

    2024

  20. [27]

    R. Wang, X. Ying, and B. Xing, `` Exploiting Temporal Correlations for 3D Human Pose Estimation ,'' IEEE Transactions on Multimedia, vol. 26, pp. 4527--4539, 2024

    2024

  21. [28]

    Z. Tang, Z. Qiu, Y. Hao, R. Hong, and T. Yao, `` 3D human pose estimation with spatio-temporal criss-cross attention,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 4790--4799

    2023

  22. [29]

    W. Li, H. Liu, H. Tang, P. Wang, and L. Van Gool, `` MHFormer: Multi-Hypothesis Transformer for 3D Human Pose Estimation ,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 13\,147--13\,156

    2022

  23. [30]

    Liu, J.-Y

    H. Liu, J.-Y. He, Z.-Q. Cheng, W. Xiang, Q. Yang, W. Chai, G. Wang, X. Bao, B. Luo, Y. Geng et al., `` PoSynDA: Multi-Hypothesis Pose Synthesis Domain Adaptation for Robust 3D Human Pose Estimation ,'' in Proceedings of the 31st ACM International Conference on Multimedia, 2023...

    2023

  24. [31]

    X. Wang, R. Girshick, A. Gupta, and K. He, ``Non-local neural networks,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2018, pp. 7794--7803

    2018

  25. [32]

    Holmquist and B

    K. Holmquist and B. Wandt, `` DiffPose: Multi-hypothesis Human Pose Estimation using Diffusion models ,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 15\,977--15\,987

    2023

  26. [33]

    J.-Y. He, X. Wu, Z.-Q. Cheng, Z. Yuan, and Y.-G. Jiang, `` DB-LSTM: Densely-connected Bi-directional LSTM for human action recognition ,'' Neurocomputing, vol. 444, pp. 319--331, 2021

    2021

  27. [34]

    W. Li, H. Liu, R. Ding, M. Liu, P. Wang, and W. Yang, ``Exploiting temporal contexts with strided transformer for 3D human pose estimation,'' IEEE Transactions on Multimedia, vol. 25, pp. 1282--1293, 2022

    2022

  28. [35]

    Zhang, Z

    J. Zhang, Z. Tu, J. Yang, Y. Chen, and J. Yuan, `` MixSTE: Seq2seq Mixed Spatio-Temporal Encoder for 3D Human Pose Estimation in Video ,'' in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 13\,232--13\,242

    2022

  29. [36]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, . Kaiser, and I. Polosukhin, ``Attention is all you need,'' Advances in Neural Information Processing Systems, vol. 30, 2017

    2017

  30. [37]

    K. Zhou, X. Han, N. Jiang, K. Jia, and J. Lu, `` HEMlets Pose: Learning Part-Centric Heatmap Triplets for Accurate 3D Human Pose Estimation ,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 2344--2353

    2019

  31. [38]

    A. Zeng, X. Sun, F. Huang, M. Liu, Q. Xu, and S. Lin, `` SRNet: Improving Generalization in 3D Human Pose Estimation with a Split-and-Recombine Approach ,'' in Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceedings, Part XIV 16. 1em...

    2020

  32. [39]

    Y. Xue, J. Chen, X. Gu, H. Ma, and H. Ma, ``Boosting monocular 3D human pose estimation with part aware attention,'' IEEE Transactions on Image Processing, vol. 31, pp. 4278--4291, 2022

    2022

  33. [40]

    Y. Wu, D. Kong, S. Wang, J. Li, and B. Yin, `` HPGCN: Hierarchical poselet-guided graph convolutional network for 3D pose estimation ,'' Neurocomputing, vol. 487, pp. 243--256, 2022

    2022

  34. [41]

    G. Hua, H. Liu, W. Li, Q. Zhang, R. Ding, and X. Xu, ``Weakly-supervised 3D human pose estimation with cross-view u-shaped graph convolutional network,'' IEEE Transactions on Multimedia, 2022

    2022

  35. [42]

    Y.-H. Wu, Y. Liu, X. Zhan, and M.-M. Cheng, `` P2T: Pyramid pooling transformer for scene understanding ,'' IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022

    2022

  36. [43]

    W. Wang, E. Xie, X. Li, D.-P. Fan, K. Song, D. Liang, T. Lu, P. Luo, and L. Shao, `` Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions ,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 568--578

    2021

  37. [44]

    Nair and G

    V. Nair and G. E. Hinton, ``Rectified linear units improve restricted boltzmann machines,'' in Proceedings of the 27th International Conference on Machine Learning (ICML-10), 2010, pp. 807--814

    2010

  38. [46]

    J. Ho, A. Jain, and P. Abbeel, ``Denoising diffusion probabilistic models,'' Advances in Neural Information Processing Systems, vol. 33, pp. 6840--6851, 2020

    2020

  39. [47]

    J. Song, C. Meng, and S. Ermon, ``Denoising diffusion implicit models,'' in International Conference on Learning Representations, 2021

    2021

  40. [48]

    Pavllo, C

    D. Pavllo, C. Feichtenhofer, D. Grangier, and M. Auli, `` 3D human pose estimation in video with temporal convolutions and semi-supervised training,'' in Conference on Computer Vision and Pattern Recognition, 2019

    2019

  41. [49]

    J. Wang, S. Yan, Y. Xiong, and D. Lin, ``Motion guided 3D pose estimation from videos,'' in European Conference on Computer Vision. 1em plus 0.5em minus 0.4em Springer, 2020, pp. 764--780

    2020

  42. [50]

    J. Zhou, T. Zhang, Z. Hayder, L. Petersson, and M. Harandi, `` Diff3DHPE: A Diffusion Model for 3D Human Pose Estimation ,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 2092--2102

    2023

  43. [51]

    W. Yang, W. Ouyang, X. Wang, J. Ren, H. Li, and X. Wang, `` 3D human pose estimation in the wild by adversarial learning,'' in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 5255--5264

    2018

  44. [52]

    Sharma, P

    S. Sharma, P. T. Varigonda, P. Bindal, A. Sharma, and A. Jain, ``Monocular 3D human pose estimation by generation and ordinal ranking,'' in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 2325--2334

    2019

  45. [53]

    T.-Y. Lin, P. Doll \'a r, R. Girshick, K. He, B. Hariharan, and S. Belongie, ``Feature pyramid networks for object detection,'' in Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2117--2125

    2017

  46. [54]

    Velickovic, G

    P. Velickovic, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y. Bengio et al., `` Graph Attention Networks ,'' International Conference on Learning Representations, 2018

    2018

  47. [55]

    K. Gong, J. Zhang, and J. Feng, `` PoseAug: A Differentiable Pose Augmentation Framework for 3D Human Pose Estimation ,'' in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    2021

  48. [56]

    X. Zhou, Q. Huang, X. Sun, X. Xue, and Y. Wei, ``Towards 3D human pose estimation in the wild: a weakly-supervised approach,'' in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 398--407

    2017

  49. [57]

    J. Lee, I. Lee, and J. Kang, ``Self-attention graph pooling,'' in Proceedings of the 36th International Conference on Machine Learning, 09--15 Jun 2019

    2019

  50. [58]

    Z. Ma, J. Xuan, Y. G. Wang, M. Li, and P. Lio, `` Path Integral Based Convolution and Pooling for Graph Neural Networks ,'' in Advances in Neural Information Processing Systems, 2020

    2020

Pith tools

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