Pith. sign in

REVIEW 3 major objections 6 minor 67 references

H$_{2}$OT: Hierarchical Hourglass Tokenizer for Efficient Video Pose Transformers

T0 review · 3 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A hierarchical hourglass tokenizer shows that full-length pose sequences are redundant: keeping only a few representative frames in the middle transformer blocks cuts FLOPs by more than half without hurting accuracy.

desk verdict Solid efficiency paper with real FLOPs gains at 50Hz, but its "accuracy preserved" headline claim breaks at 25Hz and below—worth reviewing with revisions, not a desk reject. read the letter →

arxiv 2509.06956 v1 pith:QRVB75S3 submitted 2025-09-08 cs.CV cs.AIcs.LG

Wenhao Li , Mengyuan Liu , Hong Liu , Pichao Wang , Shijian Lu , Nicu Sebe This is my paper
classification cs.CVcs.AIcs.LG
keywords 3DhumanposeestimationvideotransformertokenpruningrecoveringtemporalredundancyefficiencyHuman3.6Mseq2seqandseq2frame
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 claims that video pose transformers do not need to process every frame at every layer. It introduces H2OT, a plug-in tokenizer that prunes pose tokens down to a few representative frames in the middle blocks and then recovers the full-length output sequence. Deployed on four existing video pose transformers, it cuts computational cost by roughly 57 to 63 percent while matching or slightly improving mean per-joint position error. The claim matters because it turns a general acceleration technique into a simple, model-agnostic retrofit: existing large pose transformers can run more than twice as cheaply on the same hardware. The argument rests on temporal redundancy in high-frame-rate video: nearby frames are similar, so a handful of frames can carry the information the transformer needs.

What carries the argument

The central object is the hourglass tokenization pattern: full-length pose tokens at the input, progressively pruned representative tokens in the middle transformer blocks, and recovered full-length tokens at the output. It is implemented by the Token Pruning Module (TPM) — with the Token Pruning Sampler (TPS), a uniform temporal sampling strategy — and the Token Recovering Module (TRM) — with the Token Recovering Interpolation (TRI), a linear interpolation that fills in unselected frames. This pair makes the pruning reversible, so the model stays compatible with seq2seq and seq2frame inference pipelines.

What would settle it

Feed H2OT and the unmodified video pose transformer the same 50Hz video sequences of fast, abrupt motion (for example martial arts or rapid direction changes) and compare full-length MPJPE; if the error gap exceeds the roughly 0.5mm seen on Human3.6M, the temporal-redundancy assumption is violated. A simpler check is to run the published setup at 12.5Hz, where the paper's own results already show H2OT falling behind.

Watch

Extended reading notes

Core claim

The central discovery is that maintaining the full pose sequence across all transformer blocks is unnecessary; a few pose tokens from representative frames yield both high efficiency and estimation accuracy. H2OT realizes this with two parameter-free modules: a Token Pruning Module that uniformly samples a smaller set of frames at several stages, and a Token Recovering Module that linearly interpolates the pruned frames back to full temporal resolution after the last block. On Human3.6M, H2OT with MixSTE reduces FLOPs from 277.25G to 118.23G (down 57.4%) and improves MPJPE from 40.9mm to 40.5mm; with MotionBERT it reduces FLOPs by 63.4% with MPJPE 39.9mm versus 39.8mm; with MotionAGFormer it

Load-bearing premise

The load-bearing premise is that human motion in high-frame-rate video is smooth enough that a few uniformly sampled frames, plus linear interpolation, reproduce the full pose sequence; when the frame rate drops or motion is abrupt, the reconstruction degrades and the accuracy benefit disappears.

Editorial extensions

If this is right

  • If the claim holds, any existing video pose transformer can be accelerated by more than 2x with negligible accuracy change by inserting H2OT, without retraining from scratch or changing the regression head.
  • The same hourglass shape — process full context early, shrink in the middle, recover at the end — provides a general recipe for temporal redundancy reduction in sequential transformers, not only for human pose.
  • Because TPS and TRI are parameter-free, the speedup is not bought with extra parameters or specialized hardware; it transfers directly to deployment on resource-constrained devices.
  • The seq2frame pipeline, which outputs only the center-frame pose, can omit the recovery module entirely and still gain the same FLOP reduction while slightly improving accuracy.
  • The method also transfers to diffusion-based pose estimators, suggesting the pruning pattern applies beyond the specific transformer backbones it was tested on.

Reading between the lines

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

  • The uniform-sampling choice is likely a deliberate robustness decision: fixed sampling keeps the recovered sequence ordered for interpolation, but an adaptive sampler that concentrates tokens near motion onsets could extend the accuracy gain to mixed-speed or fast-motion videos.
  • Because the recovered frames are interpolated, downstream tasks that consume per-frame velocity or joint-angle trajectories may see artifacts invisible to MPJPE; evaluating with temporal consistency metrics would expose whether smoothing is needed.
  • The method's dependence on temporal smoothness implies a testable regime boundary: at 12.5Hz the paper's own results show accuracy degrades, so a practical deployment should gate H2OT on estimated motion speed or frame rate.
  • The hourglass design suggests that a large temporal receptive field matters more than full-resolution processing in intermediate layers; architecture search for pose transformers could exploit this by explicitly shrinking the temporal dimension in deep blocks.
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 H2OT, a hierarchical plug-and-play pruning-and-recovering framework for accelerating transformer-based 3D human pose estimation from video. It progressively prunes pose tokens in intermediate transformer blocks (TPM) using one of four strategies (cluster, attention, motion, uniform sampling) and restores full-length output tokens (TRM) via cross-attention or linear interpolation, with the default being the parameter-free TPS+TRI combination. The framework is inserted into MHFormer, MixSTE, MotionBERT, and MotionAGFormer on both seq2seq and seq2frame pipelines, reporting 52-63% FLOPs reductions with roughly constant MPJPE at 50Hz. The paper is positioned as an extension of the authors' CVPR 2024 HoT work, adding hierarchical pruning and cheaper pruning/recovering modules.

Significance. If the reported results hold, this is a practical and general-purpose efficiency tool: the FLOPs reductions are large and consistent across four architectures and two datasets, the default TPS+TRI modules are parameter-free, the comparison against input-downsampling baselines (Table 5) is the right control, and code is released. The paper also honestly acknowledges the low-FPS limitation in Sec. 4.4. However, the scientific claim is conditional on temporal smoothness at high frame rates, and the accuracy differences (0.1-0.5mm) are not statistically substantiated. The contribution over the conference version [21] is incremental (hierarchical pruning and parameter-free modules), which limits novelty but does not undermine the engineering value.

major comments (3)
  1. [Abstract; Sec. 4.4, Table 9] The abstract and conclusion claim that a few pose tokens of representative frames achieve both high efficiency and estimation accuracy. The accuracy half of this claim is established only in the 50Hz regime. Table 9 shows that H2OT w. MixSTE is already worse than MixSTE at 25Hz (40.2 vs 39.6) and the gap grows to 1.6mm at 12.5Hz (40.8 vs 39.2). This is not an experimental artifact but a consequence of the default design: TPS (Sec. 3.1.4) uniformly samples frames and TRI (Sec. 3.2.2) linearly interpolates the 3D poses, so accuracy is preserved only when motion is temporally smooth. The Sec. 4.4 discussion acknowledges the low-FPS degradation, but the word 'slightly' understates the 12.5Hz gap, and the abstract/conclusion are unqualified. Please qualify the claim to high-frame-rate input or provide a motion-adaptive sampling mechanism for low frame rates.
  2. [Tables 3 and 7] The efficiency results are robust, but the accuracy comparisons rest on single-run differences of 0.1-0.5mm without variance or multiple seeds. For example, Table 7 shows H2OT w. MotionBERT is 0.1mm worse than the re-implemented baseline (39.9 vs 39.8), H2OT w. MotionAGFormer is 0.1mm better (38.5 vs 38.4), and the 0.5mm improvement over MixSTE is used to claim superior performance. These differences are within the typical run-to-run spread of MPJPE, so the only defensible claim is accuracy parity, not improvement. Please report standard deviations over at least three seeds or soften the wording in the contribution list and Sec. 4.4.
  3. [Sec. 4.3, Tables 4 and 6] The accuracy-parity claim is sensitive to the free hyperparameters (r, b, f), and the reported optimum is selected on the test set. In Table 4, MPJPE ranges from 40.5 (r=[121,81]) to 41.3 (r=[121,27]) and 41.6 (r=[121,81,27]), a spread of 0.8-1.1mm that exceeds several claimed gains over baselines. Table 6 shows a 0.8mm spread between f=9 and f=243, described as 'almost unchanged'. These free parameters (listed in Sec. 3 and tuned per benchmark) make the central claim configuration-dependent. Please select configurations on a validation split or report a robustness curve over pruning ratios so the claim does not rely on test-set tuning.
minor comments (6)
  1. [Sec. 2] The sentence 'we are the first to perform token pruning in VPTs for model acceleration' conflicts with the paper's own prior work [21], which already prunes pose tokens in VPTs for the same purpose. Rephrase to refer to the line of work or to the specific hierarchical/parameter-free aspects.
  2. [Tables 2 and 3] The FPS conventions are confusing: Table 3 reports 9982 FPS for MixSTE seq2seq, while Table 2 reports 41 FPS/Frame for the same model; the two differ by the factor F=243. Define in the captions whether FPS is output-frame throughput or sequence throughput.
  3. [Sec. 3.1.1] The DPC-kNN description does not specify the neighborhood size k used in Eq. (1), nor how non-center tokens are assigned after the top r_m centers are chosen. Please provide these details for reproducibility.
  4. [Sec. 3.1.2] TPA selects tokens using the attention score at the b_m-th block. It is not specified whether the scores are averaged over attention heads or taken from a particular head/layer. Please specify.
  5. [Table 6 vs Sec. 3.3.3] Sec. 3.3.3 states that TRM is unnecessary for seq2frame, yet Table 6 is an ablation on the number of recovered tokens f under the seq2frame pipeline. Clarify what f controls in this setting and how the center-frame prediction is obtained when the center token is pruned.
  6. [Table 1 and Fig. 1] Minor wording issues: 'MotinAGFormer' in Table 1 should be 'MotionAGFormer'; the Fig. 1 caption describes the H2OT shape as 'trophy-shaped (pyramidal)', which is unusual; consider just 'pyramidal'. Also, the units of the frame-noise metric in Table 2 (MPJPE of 2D poses) should be stated explicitly.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the 'few tokens suffice' claim is an empirical result from ablations and external benchmarks; self-citations are transparent prior-work baselines, not load-bearing.

full rationale

The paper makes no formal derivation from assumptions to predictions; it is an empirical efficiency/accuracy study. The default combination TPS+TRI embeds the assumption that human motion at 50 Hz is temporally redundant (Sec. 3.1.4), but the accuracy claim is established by measurements (Tables 2, 3, 4, 7), not by construction. In TRI (Sec. 3.2.2), the recovered frames' poses are indeed defined by linear interpolation of the selected frames' outputs, so the full-length output is not independently predicted for every frame; however, the paper's claim is that MPJPE remains competitive, which is a testable empirical fact and is reported (e.g., Table 3: 40.5 mm vs. 40.9 mm). Table 9 shows the advantage disappears at 25 Hz and reverses at 12.5 Hz, limiting generality but not making the argument circular. Self-citation [21] appears as the previous conference version and as a baseline (Sec. 1, Tables 3, 5, 7); it is not used as evidence for the validity of the current method. The only notable methodological concern is that r and b are selected by test-set MPJPE in Table 4, which is a selection-bias/generalization issue, not circularity. No equation reduces to its own input; no fitted parameter is renamed as a prediction.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central result is empirical and rests on the smoothness/redundancy of pose video and on the sufficiency of simple sampling and interpolation. The method adds essentially no new learnable components in its final configuration (TPS+TRI), so the parameter burden is inherited from the baselines; the only added design degrees of freedom are the pruning ratios r and block indices b, tuned per model.

free parameters (4)
  • Representative token counts r per pruning stage = [121,81] for MixSTE/MotionBERT/MotionAGFormer; [175,117] for MHFormer
    Selected as 'optimal' from Table 4 ablation on Human3.6M test subjects.
  • Pruning block indices b = [0,3] for MixSTE; [0,1] for MotionBERT; [0,7] for MotionAGFormer; [0,1] for MHFormer
    Selected by test-set ablation; defines where pruning is inserted.
  • Input frame count F = 243 (MixSTE/MotionBERT/MotionAGFormer), 351 (MHFormer)
    Inherited from baseline models; not tuned by this paper.
  • Number of recovered tokens f in seq2frame = 243
    Ablation in Table 6 shows performance insensitive to f; chosen for efficiency.
assumptions (5)
  • domain assumption Adjacent video frames at 50Hz are highly redundant because poses change slowly.
    Sec 1: 'adjacent frames in a video sequence contain redundant information due to the similarity of nearby poses (50 Hz cameras used in Human3.6M)'.
  • domain assumption Uniformly sampling frames (TPS) keeps representative tokens with sufficient information.
    Sec 3.1.4: 'such a simple sampling strategy is reasonable due to data redundancy'.
  • domain assumption Linear interpolation (TRI) reconstructs full-length 3D pose sequences from representative frames without significant error.
    Sec 3.2.2; Table 9 shows this degrades at low FPS, confirming the assumption is not always satisfied.
  • domain assumption Token pruning in deep transformer blocks is a safe way to reduce compute (borrowed from image ViT literature).
    Sec 2, cited works [25], [26], [43], [44].
  • domain assumption Baseline VPT training configurations (losses, augmentation, detectors) are adopted as-is.
    Sec 4.2 implementation details inherit hyperparameters from [14], [15], [16], [27].

how reviews work

0 comments
Cite this review

Pith. "Pith review of H$_{2}$OT: Hierarchical Hourglass Tokenizer for Efficient Video Pose Transformers." pith.science (2026). https://pith.science/paper/QRVB75S3

@misc{pith2026250906956,
  author       = {Pith},
  title        = {Pith review of: H$_2$OT: Hierarchical Hourglass Tokenizer for Efficient Video Pose Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QRVB75S3}},
  note         = {Machine review of arXiv:2509.06956}
}
abstract

Transformers have been successfully applied in the field of video-based 3D human pose estimation. However, the high computational costs of these video pose transformers (VPTs) make them impractical on resource-constrained devices. In this paper, we present a hierarchical plug-and-play pruning-and-recovering framework, called Hierarchical Hourglass Tokenizer (H$_{2}$OT), for efficient transformer-based 3D human pose estimation from videos. H$_{2}$OT begins with progressively pruning pose tokens of redundant frames and ends with recovering full-length sequences, resulting in a few pose tokens in the intermediate transformer blocks and thus improving the model efficiency. It works with two key modules, namely, a Token Pruning Module (TPM) and a Token Recovering Module (TRM). TPM dynamically selects a few representative tokens to eliminate the redundancy of video frames, while TRM restores the detailed spatio-temporal information based on the selected tokens, thereby expanding the network output to the original full-length temporal resolution for fast inference. Our method is general-purpose: it can be easily incorporated into common VPT models on both seq2seq and seq2frame pipelines while effectively accommodating different token pruning and recovery strategies. In addition, our H$_{2}$OT reveals that maintaining the full pose sequence is unnecessary, and a few pose tokens of representative frames can achieve both high efficiency and estimation accuracy. Extensive experiments on multiple benchmark datasets demonstrate both the effectiveness and efficiency of the proposed method. Code and models are available at https://github.com/NationalGAILab/HoT.

Figures

Figures reproduced from arXiv: 2509.06956 by the authors.

Figure 1
Figure 1. Illustration of different VTP frameworks. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed Hierarchical Hourglass Tokenizer (H [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Illustration of our Token Pruning Cluster (TPC) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Illustration of our Token Recovering Attention (TRA). [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: Illustration of our framework on seq2frame pipeline. The pose tokens are fed into TPM to select representative tokens. After the regression head, the 3D pose of the center frame is selected as the output for evaluation. O [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Statistics visualization of selected tokens for different token pruning strategies. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results of our method on challenging in-the-wild videos. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Failure cases in challenging scenarios [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]
Figure 10
Figure 10. Figure 10: Visualization of input images, estimated 3D poses (cyan), and ground truth 3D poses (black) from three video [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 62 canonical work pages

  1. [21]

    Hourglass tok- enizer for efficient transformer-based 3D human pose estimation,

    W. Li, M. Liu, H. Liu, P . Wang, J. Cai, and N. Sebe, “Hourglass tok- enizer for efficient transformer-based 3D human pose estimation,” inCVPR, 2024, pp. 604–613

    work page 2024

  2. [1]

    Enhanced skeleton visualization for view invariant human action recognition,

    M. Liu, H. Liu, and C. Chen, “Enhanced skeleton visualization for view invariant human action recognition,”PR, vol. 68, pp. 346–362, 2017

    work page 2017

  3. [2]

    BlockGCN: Redefine topology awareness for skeleton-based action recognition,

    Y. Zhou, X. Yan, Z.-Q. Cheng, Y. Yan, Q. Dai, and X.-S. Hua, “BlockGCN: Redefine topology awareness for skeleton-based action recognition,” inCVPR, 2024, pp. 2049–2058

    work page 2024

  4. [3]

    Multi-task deep learning for real-time 3D human pose estimation and action recognition,

    D. C. Luvizon, D. Picard, and H. Tabia, “Multi-task deep learning for real-time 3D human pose estimation and action recognition,” IEEE TP AMI, vol. 43, no. 8, pp. 2752–2764, 2020

    work page 2020

  5. [4]

    3D human pose estimation in rgbd images for robotic task learning,

    C. Zimmermann, T. Welschehold, C. Dornhege, W. Burgard, and T. Brox, “3D human pose estimation in rgbd images for robotic task learning,” inICRA, 2018, pp. 1986–1992

    work page 2018

  6. [5]

    Hu- man 3D pose estimation with a tilting camera for social mobile robot interaction,

    M. Garcia-Salguero, J. Gonzalez-Jimenez, and F.-A. Moreno, “Hu- man 3D pose estimation with a tilting camera for social mobile robot interaction,”Sensors, vol. 19, no. 22, p. 4943, 2019

    work page 2019

  7. [6]

    Human-aware vision-and- language navigation: Bridging simulation to reality with dynamic human interactions,

    H. Li, M. Li, Z.-Q. Cheng, Y. Dong, Y. Zhou, J.-Y. He, Q. Dai, T. Mitamura, and A. Hauptmann, “Human-aware vision-and- language navigation: Bridging simulation to reality with dynamic human interactions,” inNeurIPS, vol. 37, 2024, pp. 119 411–119 442

    work page 2024

  8. [7]

    VNect: Real-time 3D human pose estimation with a single rgb camera,

    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 TOG, vol. 36, no. 4, pp. 1–14, 2017

    work page 2017

Show all 67 references
  1. [8]

    MotionEditor: Editing video motion via content-aware diffusion,

    S. Tu, Q. Dai, Z.-Q. Cheng, H. Hu, X. Han, Z. Wu, and Y.-G. Jiang, “MotionEditor: Editing video motion via content-aware diffusion,” inCVPR, 2024, pp. 7882–7891

    2024

  2. [9]

    Co-evolution of pose and mesh for 3D human body estimation from video,

    Y. You, H. Liu, T. Wang, W. Li, R. Ding, and X. Li, “Co-evolution of pose and mesh for 3D human body estimation from video,” in ICCV, 2023, pp. 14 963–14 973

    2023

  3. [10]

    HDFormer: High-order directed transformer for 3D human pose estimation,

    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,” inIJCAI, 2023, pp. 581–589

    2023

  4. [11]

    PoSynDA: Multi-hypothesis pose synthesis domain adaptation for robust 3D human pose estimation,

    H. Liu, J.-Y. He, Z.-Q. Cheng, W. Xiang, Q. Yang, W. Chai, G. Wang, X. Bao, B. Luo, Y. Genget al., “PoSynDA: Multi-hypothesis pose synthesis domain adaptation for robust 3D human pose estimation,” inACM MM, 2023, pp. 5542–5551

    2023

  5. [12]

    APP: Adaptive pose pooling for 3D human pose estimation from videos,

    J. Zhang, M. Liu, H. Liu, G. Wang, and W. Li, “APP: Adaptive pose pooling for 3D human pose estimation from videos,” inACM MM, 2024, pp. 1672–1681

    2024

  6. [13]

    3D human pose estimation with spatial and temporal transformers,

    C. Zheng, S. Zhu, M. Mendieta, T. Yang, C. Chen, and Z. Ding, “3D human pose estimation with spatial and temporal transformers,” inICCV, 2021, pp. 11 656–11 665

    2021

  7. [14]

    MHFormer: Multi-hypothesis transformer for 3D human pose estimation,

    W. Li, H. Liu, H. Tang, P . Wang, and L. Van Gool, “MHFormer: Multi-hypothesis transformer for 3D human pose estimation,” in CVPR, 2022, pp. 13 147–13 156

    2022

  8. [15]

    MixSTE: Seq2seq mixed spatio-temporal encoder for 3D human pose estimation in video,

    J. Zhang, Z. Tu, J. Yang, Y. Chen, and J. Yuan, “MixSTE: Seq2seq mixed spatio-temporal encoder for 3D human pose estimation in video,” inCVPR, 2022, pp. 13 232–13 242

    2022

  9. [16]

    MotionBERT: A unified perspective on learning human motion representations,

    W. Zhu, X. Ma, Z. Liu, L. Liu, W. Wu, and Y. Wang, “MotionBERT: A unified perspective on learning human motion representations,” inICCV, 2023, pp. 15 085–15 099

    2023

  10. [17]

    TCPFormer: Learning temporal correlation with implicit pose proxy for 3D human pose estimation,

    J. Liu, M. Liu, H. Liu, and W. Li, “TCPFormer: Learning temporal correlation with implicit pose proxy for 3D human pose estimation,” inAAAI, 2025

    2025

  11. [18]

    P-STMO: Pre-trained spatial temporal many-to-one model for 3D human pose estimation,

    W. Shan, Z. Liu, X. Zhang, S. Wang, S. Ma, and W. Gao, “P-STMO: Pre-trained spatial temporal many-to-one model for 3D human pose estimation,” inECCV, 2022

    2022

  12. [19]

    Exploiting temporal contexts with strided transformer for 3D human pose estimation,

    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 TMM, vol. 25, pp. 1282–1293, 2022

    2022

  13. [20]

    Uplift and upsample: Efficient 3D human pose estimation with uplifting transformers,

    M. Einfalt, K. Ludwig, and R. Lienhart, “Uplift and upsample: Efficient 3D human pose estimation with uplifting transformers,” inWACV, 2023, pp. 2903–2913

    2023

  14. [22]

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

    D. Pavllo, C. Feichtenhofer, D. Grangier, and M. Auli, “3D human pose estimation in video with temporal convolutions and semi- supervised training,” inCVPR, 2019, pp. 7753–7762

    2019

  15. [23]

    Attention mechanism exploits temporal contexts: Real-time 3D human pose reconstruction,

    R. Liu, J. Shen, H. Wang, C. Chen, S.-c. Cheung, and V . Asari, “Attention mechanism exploits temporal contexts: Real-time 3D human pose reconstruction,” inCVPR, 2020, pp. 5064–5073

    2020

  16. [24]

    Hu- man3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments,

    C. Ionescu, D. Papava, V . Olaru, and C. Sminchisescu, “Hu- man3.6M: Large scale datasets and predictive methods for 3D human sensing in natural environments,”IEEE TP AMI, vol. 36, no. 7, pp. 1325–1339, 2013

    2013

  17. [25]

    DynamicViT: Efficient vision transformers with dynamic token sparsification,

    Y. Rao, W. Zhao, B. Liu, J. Lu, J. Zhou, and C.-J. Hsieh, “DynamicViT: Efficient vision transformers with dynamic token sparsification,” inNeurIPS, vol. 34, 2021, pp. 13 937–13 949

    2021

  18. [26]

    VTC-LFC: Vision transformer compression with low-frequency components,

    Z. Wang, H. Luo, P . Wang, F. Ding, F. Wang, and H. Li, “VTC-LFC: Vision transformer compression with low-frequency components,” inNeurIPS, vol. 35, 2022, pp. 13 974–13 988

    2022

  19. [27]

    MotionAGFormer: Enhancing 3D human pose estimation with a transformer-gcnformer network,

    S. Mehraban, V . Adeli, and B. Taati, “MotionAGFormer: Enhancing 3D human pose estimation with a transformer-gcnformer network,” inWACV, 2024, pp. 6920–6930

    2024

  20. [28]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” inNeurIPS, 2017, pp. 5998–6008

    2017

  21. [29]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inICLR, 2021

    2021

  22. [30]

    Swin Transformer: Hierarchical vision transformer using shifted windows,

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, and B. Guo, “Swin Transformer: Hierarchical vision transformer using shifted windows,” inICCV, 2021, pp. 10 012–10 022

    2021

  23. [31]

    GSRFormer: Grounded situation recognition transformer with alternate semantic attention refinement,

    Z.-Q. Cheng, Q. Dai, S. Li, T. Mitamura, and A. Hauptmann, “GSRFormer: Grounded situation recognition transformer with alternate semantic attention refinement,” inACM MM, 2022, pp. 3272–3281

    2022

  24. [32]

    Implicit temporal modeling with learnable alignment for video recognition,

    S. Tu, Q. Dai, Z. Wu, Z.-Q. Cheng, H. Hu, and Y.-G. Jiang, “Implicit temporal modeling with learnable alignment for video recognition,” inICCV, 2023, pp. 19 936–19 947

    2023

  25. [33]

    Adaptive multi-view and temporal fusing transformer for 3D human pose estimation,

    H. Shuai, L. Wu, and Q. Liu, “Adaptive multi-view and temporal fusing transformer for 3D human pose estimation,”IEEE TP AMI, vol. 45, no. 4, pp. 4122–4135, 2022. IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 14

    2022

  26. [34]

    SVTformer: Spatial-view- temporal transformer for multi-view 3d human pose estimation,

    W. Zhang, M. Liu, H. Liu, and W. Li, “SVTformer: Spatial-view- temporal transformer for multi-view 3d human pose estimation,” inAAAI, vol. 39, no. 10, 2025, pp. 10 148–10 156

    2025

  27. [35]

    A simple yet effective baseline for 3D human pose estimation,

    J. Martinez, R. Hossain, J. Romero, and J. J. Little, “A simple yet effective baseline for 3D human pose estimation,” inICCV, 2017, pp. 2640–2649

    2017

  28. [36]

    MobileHumanPose: Toward real- time 3D human pose estimation in mobile devices,

    S. Choi, S. Choi, and C. Kim, “MobileHumanPose: Toward real- time 3D human pose estimation in mobile devices,” inCVPR, 2021, pp. 2328–2338

    2021

  29. [37]

    DeciWatch: A simple baseline for 10x efficient 2D and 3D pose estimation,

    A. Zeng, X. Ju, L. Yang, R. Gao, X. Zhu, B. Dai, and Q. Xu, “DeciWatch: A simple baseline for 10x efficient 2D and 3D pose estimation,” inECCV, 2022, pp. 607–624

    2022

  30. [38]

    PoseFormerV2: Exploring frequency domain for efficient and robust 3D human pose estimation,

    Q. Zhao, C. Zheng, M. Liu, P . Wang, and C. Chen, “PoseFormerV2: Exploring frequency domain for efficient and robust 3D human pose estimation,” inCVPR, 2023, pp. 8877–8886

    2023

  31. [39]

    MixSynthFormer: A transformer encoder-like structure with mixed synthetic self-attention for efficient human pose estimation,

    Y. Sun, A. W. Dougherty, Z. Zhang, Y. K. Choi, and C. Wu, “MixSynthFormer: A transformer encoder-like structure with mixed synthetic self-attention for efficient human pose estimation,” in ICCV, 2023, pp. 14 884–14 893

    2023

  32. [40]

    Learn- ing to reconstruct 3D human pose and shape via model-fitting in the loop,

    N. Kolotouros, G. Pavlakos, M. J. Black, and K. Daniilidis, “Learn- ing to reconstruct 3D human pose and shape via model-fitting in the loop,” inICCV, 2019, pp. 2252–2261

    2019

  33. [41]

    Exemplar fine-tuning for 3D human model fitting towards in-the-wild 3D human pose estimation,

    H. Joo, N. Neverova, and A. Vedaldi, “Exemplar fine-tuning for 3D human model fitting towards in-the-wild 3D human pose estimation,” in3DV, 2021, pp. 42–52

    2021

  34. [42]

    PARE: Part attention regressor for 3D human body estimation,

    M. Kocabas, C.-H. P . Huang, O. Hilliges, and M. J. Black, “PARE: Part attention regressor for 3D human body estimation,” inICCV, 2021, pp. 11 127–11 137

    2021

  35. [43]

    TORE: Token reduction for efficient human mesh recovery with transformer,

    Z. Dou, Q. Wu, C. Lin, Z. Cao, Q. Wu, W. Wan, T. Komura, and W. Wang, “TORE: Token reduction for efficient human mesh recovery with transformer,” inICCV, 2023, pp. 15 143–15 155

    2023

  36. [44]

    Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers,

    S. Long, Z. Zhao, J. Pi, S. Wang, and J. Wang, “Beyond attentive tokens: Incorporating token importance and diversity for efficient vision transformers,” inCVPR, 2023, pp. 10 334–10 343

    2023

  37. [45]

    SPViT: Enabling faster vision transformers via latency-aware soft token pruning,

    Z. Kong, P . Dong, X. Ma, X. Meng, W. Niu, M. Sun, X. Shen, G. Yuan, B. Ren, H. Tanget al., “SPViT: Enabling faster vision transformers via latency-aware soft token pruning,” inECCV, 2022, pp. 620–640

    2022

  38. [46]

    Making vision transformers efficient from a token sparsification view,

    S. Chang, P . Wang, M. Lin, F. Wang, D. J. Zhang, R. Jin, and M. Z. Shou, “Making vision transformers efficient from a token sparsification view,” inCVPR, 2023, pp. 6195–6205

    2023

  39. [47]

    PPT: Token-pruned pose transformer for monocular and multi-view human pose estimation,

    H. Ma, Z. Wang, Y. Chen, D. Kong, L. Chen, X. Liu, X. Yan, H. Tang, and X. Xie, “PPT: Token-pruned pose transformer for monocular and multi-view human pose estimation,” inECCV, 2022, pp. 424– 442

    2022

  40. [48]

    TCFormer: Visual recognition via token clustering transformer,

    W. Zeng, S. Jin, L. Xu, W. Liu, C. Qian, W. Ouyang, P . Luo, and X. Wang, “TCFormer: Visual recognition via token clustering transformer,”IEEE TP AMI, 2024

    2024

  41. [49]

    GTPT: Group-based token pruning transformer for efficient human pose estimation,

    H. Wang, J. Liu, J. Tang, G. Wu, B. Xu, Y. Chou, and Y. Wang, “GTPT: Group-based token pruning transformer for efficient human pose estimation,” inECCV, 2024

    2024

  42. [50]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” inCVPR, 2016, pp. 770–778

    2016

  43. [51]

    Very deep convolutional networks for large-scale image recognition,

    K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” inICLR, 2015

    2015

  44. [52]

    Study on density peaks clustering based on k-nearest neighbors and principal component analysis,

    M. Du, S. Ding, and H. Jia, “Study on density peaks clustering based on k-nearest neighbors and principal component analysis,” Knowledge-Based Systems, vol. 99, pp. 135–145, 2016

    2016

  45. [53]

    A-ViT: Adaptive tokens for efficient vision trans- former,

    H. Yin, A. Vahdat, J. M. Alvarez, A. Mallya, J. Kautz, and P . Molchanov, “A-ViT: Adaptive tokens for efficient vision trans- former,” inCVPR, 2022, pp. 10 809–10 818

    2022

  46. [54]

    Token pooling in vision transformers for image classification,

    D. Marin, J.-H. R. Chang, A. Ranjan, A. Prabhu, M. Rastegari, and O. Tuzel, “Token pooling in vision transformers for image classification,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2023, pp. 12–21

    2023

  47. [55]

    Not all patches are what you need: Expediting vision transformers via token reorganizations,

    Y. Liang, C. Ge, Z. Tong, Y. Song, J. Wang, and P . Xie, “Not all patches are what you need: Expediting vision transformers via token reorganizations,” inICLR, 2022

    2022

  48. [56]

    DiffRate: Differentiable compression rate for efficient vision transformers,

    M. Chen, W. Shao, P . Xu, M. Lin, K. Zhang, F. Chao, R. Ji, Y. Qiao, and P . Luo, “DiffRate: Differentiable compression rate for efficient vision transformers,” inICCV, 2023, pp. 17 164–17 174

    2023

  49. [57]

    Multi-hypothesis represen- tation learning for transformer-based 3D human pose estimation,

    W. Li, H. Liu, H. Tang, and P . Wang, “Multi-hypothesis represen- tation learning for transformer-based 3D human pose estimation,” PR, vol. 141, p. 109631, 2023

    2023

  50. [58]

    Monocular 3D human pose estimation in the wild using improved CNN supervision,

    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,” in3DV, 2017, pp. 506–516

    2017

  51. [59]

    GraphMLP: A graph MLP-like architecture for 3D human pose estimation,

    W. Li, M. Liu, H. Liu, T. Guo, T. Wang, H. Tang, and N. Sebe, “GraphMLP: A graph MLP-like architecture for 3D human pose estimation,”PR, vol. 158, p. 110925, 2025

    2025

  52. [60]

    Locally connected network for monocular 3D human pose estimation,

    H. Ci, X. Ma, C. Wang, and Y. Wang, “Locally connected network for monocular 3D human pose estimation,”IEEE TP AMI, vol. 44, no. 3, pp. 1429–1442, 2020

    2020

  53. [61]

    Monocular 3D pose estimation via pose grammar and data augmentation,

    Y. Xu, W. Wang, T. Liu, X. Liu, J. Xie, and S.-C. Zhu, “Monocular 3D pose estimation via pose grammar and data augmentation,”IEEE TP AMI, vol. 44, no. 10, pp. 6327–6344, 2021

    2021

  54. [62]

    Cascaded deep monocular 3D human pose estimation with evolutionary training data,

    S. Li, L. Ke, K. Pratama, Y.-W. Tai, C.-K. Tang, and K.-T. Cheng, “Cascaded deep monocular 3D human pose estimation with evolutionary training data,” inCVPR, 2020, pp. 6173–6183

    2020

  55. [63]

    Stacked hourglass networks for human pose estimation,

    A. Newell, K. Yang, and J. Deng, “Stacked hourglass networks for human pose estimation,” inECCV, 2016, pp. 483–499

    2016

  56. [64]

    3D human pose estimation with spatio-temporal criss-cross attention,

    Z. Tang, Z. Qiu, Y. Hao, R. Hong, and T. Yao, “3D human pose estimation with spatio-temporal criss-cross attention,” inCVPR, 2023, pp. 4790–4799

    2023

  57. [65]

    KTPFormer: Kinematics and trajectory prior knowledge-enhanced transformer for 3D human pose estimation,

    J. Peng, Y. Zhou, and P . Mok, “KTPFormer: Kinematics and trajectory prior knowledge-enhanced transformer for 3D human pose estimation,” inCVPR, 2024, pp. 1123–1132

    2024

  58. [66]

    Diffusion-based 3D human pose estimation with multi-hypothesis aggregation,

    W. Shan, Z. Liu, X. Zhang, Z. Wang, K. Han, S. Wang, S. Ma, and W. Gao, “Diffusion-based 3D human pose estimation with multi-hypothesis aggregation,” inICCV, 2023, pp. 14 761–14 771

    2023

  59. [67]

    DiffPose: Toward more reliable 3D pose estimation,

    J. Gong, L. G. Foo, Z. Fan, Q. Ke, H. Rahmani, and J. Liu, “DiffPose: Toward more reliable 3D pose estimation,” inCVPR, 2023, pp. 13 041–13 051. Wenhao Liis currently a Postdoctoral Re- searcher at the College of Computing and Data Science, Nanyang Technological University, Si...

    2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.