SpatioTemporal Learning for Human Pose Estimation in Sparsely-Labeled Videos

Shaojing Fan; Sifan Wu; Yingying Jiao; Zhenguang Liu; Zheqi Wu; Zhigang Wang; Zhuoyue Xu

REVIEW 3 major objections 7 minor 47 references

SpatioTemporal Learning for Human Pose Estimation in Sparsely-Labeled Videos

T0 review · 3 major / 7 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read STDPose claims that jointly modeling multi-frame visual features and pose heatmaps with a Dynamic-Aware Mask sets new state-of-the-art results for video pose propagation and estimation, reducing the need for dense manual annotation.

desk verdict Solid engineering with a real backbone confound and an un-implementable MI term; the pseudo-label result is the best part. read the letter →

arxiv 2501.15073 v1 pith:DVUDIUAC submitted 2025-01-25 cs.CV

Yingying Jiao , Zhigang Wang , Sifan Wu , Shaojing Fan , Zhenguang Liu , Zhuoyue Xu , Zheqi Wu This is my paper · ORCID

classification cs.CV

keywords videoposeestimationpropagationsparsely-labeledvideosdynamic-awaremaskvisiontransformercross-attentionmutualinformationTrack

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

STDPose is a framework for estimating human poses in videos where only a fraction of frames carry annotations. The paper argues that existing video pose models fail to exploit long-range temporal dependencies and treat pose heatmaps and visual features separately, so it introduces a transformer-based encoder that fuses both and a Dynamic-Aware Mask that highlights motion regions. The authors report state-of-the-art mAP on PoseTrack2017, PoseTrack2018, and PoseTrack2021 for both pose propagation and pose estimation. They also show that pseudo-labels from propagation let the estimator reach 85.2 mAP with half the frames labeled and 84.3 mAP with about a quarter, suggesting that dense annotation may be unnecessary.

What carries the argument

The load-bearing mechanism is the Dynamic-Aware Mask together with the cross-attention aggregation in STDA. The mask is computed from heatmap residuals between auxiliary and key frames, passed through a modified sigmoid $\text{Sigm}(x,k,\theta)=1/(1+e^{-k(|x|-\theta)})$ that responds equally to positive and negative residuals, then compressed and softmax-normalized; this highlights where joints actually moved. STDA multiplies the merged heatmaps $\tilde H_t$ by this mask, concatenates the result with $\tilde H_t$, embeds it into tokens, and applies self- and cross-attention against the fused visual feature $\tilde F_t$. The mutual-information objective, approximated as $I(y_t;\tilde F_t)-I(\tilde F_t;F_t)$ and similarly for heatmaps, supervises that the fused representations add label-relevant information beyond the key-frame features.

What would settle it

Train the strongest baselines (PoseWarper, DCPose, FAMI-Pose) with the same ViT backbone and input resolution, or train STDPose with HRNet; if the mAP gaps on PoseTrack2017 (2.2, 1.0, and so on) disappear or reverse, the central SOTA claim is largely a backbone artifact.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that coupling temporal heatmap synthesis with motion-aware attention improves sparse-label video pose estimation. The SpatioTemporal Representation Encoder fuses features from key and auxiliary frames with transformer blocks (TFF) and merges per-joint heatmaps over time with convolutions (TKS), while the Dynamic-Aware Mask is built from forward and backward heatmap residuals. The STDA module then lets the masked heatmaps attend to the fused features via cross-attention, and a mutual-information loss filters task-irrelevant temporal information. With this design, STDPose reports 90.9 mAP for pose propagation and 87.4 mAP for pose estimation on PoseTrack2017, with the largest gains on wrists and ankles, and claims competitive sparse-label performance.

Load-bearing premise

The reported gains are credited to the proposed modules, but every comparison uses a COCO-pretrained Vision Transformer for STDPose and HRNet for the baselines, with no backbone-controlled experiment, so the modules' contribution is not isolated.

Editorial extensions

If this is right

If the reported results hold, a pose model can be trained from every 7th labeled frame and still reach 90.9 mAP on PoseTrack2017, so annotation cost drops sharply.
Pseudo-labels produced by propagation give 84.3 mAP with 26.7% manually labeled frames, close to the 84.8 mAP of a fully labeled baseline, meaning dense manual annotation may become unnecessary.
The largest improvements on wrist and ankle joints suggest the motion-aware mask helps exactly the joints most affected by blur and occlusion.
Replacing deformable convolutions with cross-attention means the model no longer relies on local receptive fields, which the paper claims addresses global spatial correlations.

Reading between the lines

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

A controlled backbone swap would separate the contributions of the modules from the ViT backbone, since all reported comparisons pair STDPose's ViT against HRNet baselines.
The Dynamic-Aware Mask could be reused as a motion-attention module in other video tasks like action recognition or object segmentation, since it only requires heatmap-like residuals.
Learning the sigmoid parameters $k$ and $\theta$ instead of fixing them might adapt the mask to videos with different motion speeds, given the paper's own ablation shows flat sensitivity.
Iterative pseudo-label self-training beyond one propagation round could push sparse-label accuracy closer to full supervision.

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.

Desk Editor's Note

read the letter

This paper is a competent engineering contribution to video pose estimation, but the headline SOTA claim has a confound the authors don't fully address. The genuinely new thing is the specific pipeline—ViT backbone, temporal feature fusion, temporal keypoint synthesis, a Dynamic-Aware Mask, cross-attention aggregation, and an MI-style loss—aimed at propagating pose labels from sparsely labeled frames. The pseudo-label experiment is the most practically useful piece: with 50% labeled frames they report 85.2 mAP, 0.4 above FAMI-Pose, and with 26.7% they're within 0.5. That's a real result, though the comparison is narrow.

The soft spot is the backbone. All SOTA comparisons in Tables 2 and 3 pair STDPose's ViT with HRNet-based baselines. The propagation ablation in Table 4(a) does include a stripped ViT+head baseline at 86.8 mAP, so the +4.1 mAP from the full model is meaningful for propagation. But Tables 2 and 3 lack that control. Gains over DiffPose and TDMI are only 0.5–1.0 mAP, which could be backbone strength. I don't think this is fatal—the ablation suggests the modules do something—but the pose estimation SOTA claim is not established without a ViT-backed baseline.

The second real gap is the mutual information loss. Appendix Eq. (11) is written as abstract MI terms, with no estimator, surrogate, or code. Dropping the conditional term is a heuristic, not a derivation. The +0.6 mAP from the MI component in Table 4(f) is not independently reproducible. That's a reproducibility gap, not a fatal one, but it should be disclosed. The k and theta hyperparameters are tuned on the same validation set, though Table 5 shows the surface is flat, so this is minor.

Overall, the paper is honest and internally consistent. It's incremental but thoroughly evaluated on real benchmarks. Someone working on reducing annotation cost will get value from the pseudo-label part and the time-interval analysis. I'd take it to our reading group. I'd send it to peer review, asking for a backbone-controlled comparison and a concrete MI loss before accepting.

Referee Report

3 major / 7 minor

Summary. The paper proposes STDPose, a transformer-based framework for video pose propagation and pose estimation from sparsely-labeled videos. STDPose consists of a SpatioTemporal Representation Encoder (STRE) with temporal feature fusion (TFF) and temporal keypoint synthesis (TKS), a Dynamic-Aware Mask (DAM), a SpatioTemporal Dynamics Aggregation (STDA) module, and a mutual-information-based loss. The authors report state-of-the-art results on three PoseTrack datasets for both pose propagation (90.9 mAP on PoseTrack2017) and pose estimation (87.4 mAP on PoseTrack2017), along with a pseudo-label training experiment that achieves 84.3 mAP with 26.7% labeled frames and 85.2 mAP with 50% labeled frames.

Significance. If the results hold, the framework would be a meaningful step toward reducing annotation cost in video pose estimation while improving accuracy under occlusion and blur. The paper provides a reasonably complete ablation study (Table 4), a parameter analysis for the modified sigmoid (Table 5), and experiments at multiple input resolutions (Table 6), and the tables are internally consistent. However, two load-bearing issues prevent verification of the central claims: the SOTA comparisons are confounded by the backbone choice (ViT vs. HRNet), and the mutual-information objective is not implementable as written. The backbone issue is particularly serious because the margins over strong baselines in Table 2 are only 0.5-1.0 mAP.

major comments (3)

[Sec. 4.2-4.3, Tables 1-3 and Appendix Table 6] The state-of-the-art comparisons are confounded by the backbone. STDPose uses a COCO-pretrained ViT at 256x192, whereas all compared methods use HRNet-based backbones, some at 384x288. Table 4(a) provides a ViT+head baseline of 86.8 mAP for the propagation task, which partially controls for the propagation claim, but no ViT-backed baseline is reported for the pose-estimation protocol in Table 2 or for the pseudo-label experiment in Table 3. Since the gains over DiffPose and TDMI in Table 2 are only 0.5-1.0 mAP, the central SOTA claim for pose estimation is not established. Please provide either an HRNet-backed version of STDPose or ViT-backed versions of the baselines for all three protocols, or explicitly qualify the claim as a backbone-plus-method result.
[Appendix A, Eqs. (8)-(11); Sec. 3.1, Eq. (2)] The mutual-information objective is not implementable as written. Eq. (11) is a difference of intractable mutual-information terms, and the paper provides no variational estimator, surrogate bound, sampling procedure, or code. The simplification in Eqs. (9) and (10) drops I(F;F|y) and I(H;H|y) based on the unverified assumption that these conditional MI terms become negligible after sufficient training. Because Table 4(f) attributes a 0.6 mAP improvement to this loss term, the reader cannot reproduce the result. Please specify a tractable estimator (e.g., a variational bound or a contrastive loss) and, ideally, release code for this component.
[Tables 1-3] The reported improvements over the strongest baselines are small, and no uncertainty quantification is provided. For example, on PoseTrack2017 pose estimation, STDPose achieves 87.4 mAP versus 86.4 for DiffPose; on PoseTrack2018, it achieves 84.5 versus 83.6 for TDMI. Without multiple runs, standard deviations, or significance tests, it is unclear whether these margins are meaningful. Please report variance across seeds or otherwise justify the stability of the results.

minor comments (7)

[Tables 3 and 4] The relationship between T and the labeled-frame ratio is not explained. If one frame is labeled every T frames, the ratio should be approximately 1/T (e.g., 14.3% for T=7 and 25% for T=4), but the tables list 16.7% and 26.7%. Please state how these percentages are computed, including any effect of the 'only 30 consecutive frames' restriction described in Appendix C.
[Sec. 4.4] The phrase 'nearly matching a 100% ideal precision of manual annotation' in the discussion of Table 4 is not supported by a ceiling experiment; no result is reported for training with full manual annotations under the same protocol. Please rephrase or provide such a baseline.
[Figure 2] Figure 2 is very dense and difficult to read; the font sizes and arrow labels are small. A cleaner, higher-level diagram would help readers understand the data flow.
[Sec. 1, first paragraph] There is a typo in the Introduction: 'spatiotemporal representatons' should be 'spatiotemporal representations.'
[Table 5 and Sec. 3.3] The parameter sensitivity of the modified sigmoid is reported, but the hyperparameters α and β in Eq. (2) are not ablated. Since α and β are set to 0.1 and 0.01 without tuning, it would be informative to show how the MI loss contribution varies with these weights, especially given that the MI term is not reproducible as written.
[Sec. 4.3] The comparison with methods that use four auxiliary frames (TDMI, DSTA) vs. STDPose's two auxiliary frames is useful, but the paper should clarify whether the cited numbers for those methods were obtained with their original four-frame setting; if so, the comparison is not fully controlled.
[Appendix A, Eq. (6)] The notation in Eq. (6) is confusing, particularly the use of H^{i,j} without defining the superscript j and the concatenation of multiple Conv outputs. Please add explicit dimensions and a clearer description of the temporal keypoint synthesis.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity: benchmark claims rest on external PoseTrack evaluations and compositional ablations; the score reflects only minor non-load-bearing self-citations and two evidence gaps that are correctness/reproducibility risks, not circular reductions.

full rationale

STDPose's headline claims are empirical SOTA results on PoseTrack2017/2018/2021. Pose propagation (Table 1), pose estimation (Table 2), and pseudo-label training (Table 3) are all evaluated against external validation sets, so the reported numbers are not derived from the method's own definitions. The ablation in Table 4 includes a stripped ViT+head baseline (86.8 mAP) and adds TFF, TKS, STDA, DAM, and MI one at a time; this is an internal contribution analysis, not a circular reduction. Pseudo-label generation is a standard self-training pipeline: a propagation model produces labels for unlabeled frames, and a separate pose-estimation model is trained and tested on held-out validation data; nothing here is statistically forced by construction. The few self-citations (Liu et al. 2022a for MI inspiration/difficulty, Wu et al. 2024a as a recent-methods example) are not load-bearing: no central claim rests on an unverified result from the authors' own prior work. Two genuine evidence gaps should be noted as correctness/reproducibility risks rather than circularity: (1) Appendix Eq. (11) writes LMI with intractable mutual-information terms and no variational estimator, so the 0.6 mAP MI gain in Table 4(f) is not independently reproducible from the paper; (2) Tables 2-3 compare STDPose (ViT backbone) with HRNet-based baselines without a backbone-matched control, so part of the SOTA margin could reflect backbone strength. Neither gap makes a 'prediction' equal to its input by construction. Overall circularity burden is low; score 2 reflects only minor non-load-bearing self-citation and the flagged evidence gaps, not a circular derivation.

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

The central claim rests on a small set of hyperparameters, an unproven simplification of the mutual information objective, and an unstated comparability assumption between ViT and HRNet backbones. No new physical entities are introduced.

free parameters (4)

alpha = 0.1
Weight for the feature mutual information term in Eq. (2), set by hand and described as not densely tuned.
beta = 0.01
Weight for the heatmap mutual information term in Eq. (2), set by hand and described as not densely tuned.
k = 1.5
Slope parameter in the modified sigmoid for the Dynamic-Aware Mask, selected by validation-set ablation in Appendix Table 5.
theta = 0.5
Threshold parameter in the modified sigmoid for the Dynamic-Aware Mask, selected by validation-set ablation in Appendix Table 5.

assumptions (4)

ad hoc to paper The dropped conditional mutual information terms I(F̃;F|y) and I(H̃;H|y) become negligible after sufficient training.
Appendix A Eqs. 9-10 remove these terms based on a heuristic claim that task-irrelevant information diminishes during optimization, without an experiment or bound supporting the claim.
domain assumption A COCO-pretrained Vision Transformer backbone is directly comparable to the HRNet backbones used by all compared methods.
Sec 4.1 and Appendix Table 6: STDPose uses ViT while prior baselines use HRNet, and no controlled experiment isolates the backbone from the proposed modules.
domain assumption An external object detector provides correct person crops for the top-down pose estimation pipeline.
Sec 3 Problem formulation: the method depends on an object detector, but no detector details, accuracy, or failure analysis are provided.
standard math Mutual information identities used for the MI objective simplification are standard.
Appendix A Eqs. 7-10 rely on the chain rule and symmetry of mutual information, which are standard mathematical facts.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SpatioTemporal Learning for Human Pose Estimation in Sparsely-Labeled Videos." pith.science (2026). https://pith.science/paper/DVUDIUAC

@misc{pith2026250115073,
  author       = {Pith},
  title        = {Pith review of: SpatioTemporal Learning for Human Pose Estimation in Sparsely-Labeled Videos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DVUDIUAC}},
  note         = {Machine review of arXiv:2501.15073}
}

read the original abstract

Human pose estimation in videos remains a challenge, largely due to the reliance on extensive manual annotation of large datasets, which is expensive and labor-intensive. Furthermore, existing approaches often struggle to capture long-range temporal dependencies and overlook the complementary relationship between temporal pose heatmaps and visual features. To address these limitations, we introduce STDPose, a novel framework that enhances human pose estimation by learning spatiotemporal dynamics in sparsely-labeled videos. STDPose incorporates two key innovations: 1) A novel Dynamic-Aware Mask to capture long-range motion context, allowing for a nuanced understanding of pose changes. 2) A system for encoding and aggregating spatiotemporal representations and motion dynamics to effectively model spatiotemporal relationships, improving the accuracy and robustness of pose estimation. STDPose establishes a new performance benchmark for both video pose propagation (i.e., propagating pose annotations from labeled frames to unlabeled frames) and pose estimation tasks, across three large-scale evaluation datasets. Additionally, utilizing pseudo-labels generated by pose propagation, STDPose achieves competitive performance with only 26.7% labeled data.

Figures

Figures reproduced from arXiv: 2501.15073 by the authors.

**Figure 2.** The overall pipeline of our STDPose framework. Given an input sequence [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Visual results of our STDPose on the PoseTrack2017 (Iqbal, Milan, and Gall 2017) dataset include challenging scenes, [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Visual results of our STDPose on the PoseTrack2018 (Andriluka et al. 2018) dataset include challenging scenes, such [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]

**Figure 5.** Figure 5: Visual results of our STDPose on the PoseTrack2021 (Doering et al. 2022) dataset include challenging scenes, such [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

47 extracted references · 31 canonical work pages

[1]

Andriluka, M.; Iqbal, U.; Insafutdinov, E.; Pishchulin, L.; Milan, A.; Gall, J.; and Schiele, B. 2018. Posetrack: A benchmark for human pose estimation and tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, 5167--5176

work page 2018
[2]

Artacho, B.; and Savakis, A. 2020. Unipose: Unified human pose estimation in single images and videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7035--7044

work page 2020
[3]

Bertasius, G.; Feichtenhofer, C.; Tran, D.; Shi, J.; and Torresani, L. 2019. Learning temporal pose estimation from sparsely-labeled videos. Advances in neural information processing systems, 32

work page 2019
[4]

Chen, X.; Wang, X.; Zhou, J.; Qiao, Y.; and Dong, C. 2023. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22367--22377

2023
[5]

Dai, J.; Qi, H.; Xiong, Y.; Li, Y.; Zhang, G.; Hu, H.; and Wei, Y. 2017. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, 764--773

2017
[6]

Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, 248--255. Ieee

2009
[7]

Doering, A.; Chen, D.; Zhang, S.; Schiele, B.; and Gall, J. 2022. Posetrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20963--20972

work page 2022
[8]

Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

arXiv 2020

Show all 47 references

[9]

Farneb \"a ck, G. 2003. Two-frame motion estimation based on polynomial expansion. In Image Analysis: 13th Scandinavian Conference, SCIA 2003 Halmstad, Sweden, June 29--July 2, 2003 Proceedings 13, 363--370. Springer

2003
[10]

Feng, R.; Gao, Y.; Ma, X.; Tse, T. H. E.; and Chang, H. J. 2023 a . Mutual information-based temporal difference learning for human pose estimation in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 17131--17141

2023
[11]

Feng, R.; Gao, Y.; Tse, T. H. E.; Ma, X.; and Chang, H. J. 2023 b . DiffPose: SpatioTemporal diffusion model for video-based human pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 14861--14872

2023
[12]

He, J.; and Yang, W. 2024. Video-Based Human Pose Regression via Decoupled Space-Time Aggregation. arXiv preprint arXiv:2403.19926

2024 arXiv
[13]

D.; Fedorov, A.; Lavoie-Marchildon, S.; Grewal, K.; Bachman, P.; Trischler, A.; and Bengio, Y

Hjelm, R. D.; Fedorov, A.; Lavoie-Marchildon, S.; Grewal, K.; Bachman, P.; Trischler, A.; and Bengio, Y. 2018. Learning deep representations by mutual information estimation and maximization. arXiv preprint arXiv:1808.06670

2018 arXiv
[14]

Ilg, E.; Mayer, N.; Saikia, T.; Keuper, M.; Dosovitskiy, A.; and Brox, T. 2017. Flownet 2.0: Evolution of optical flow estimation with deep networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, 2462--2470

2017
[15]

Iqbal, U.; Milan, A.; and Gall, J. 2017. Posetrack: Joint multi-person pose estimation and tracking. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2011--2020

2017
[16]

Jin, K.-M.; Lee, G.-H.; and Lee, S.-W. 2022. OTPose: occlusion-aware transformer for pose estimation in sparsely-labeled videos. In 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC), 3255--3260. IEEE

2022
[17]

C.; Lo, W.-Y.; et al

Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A. C.; Lo, W.-Y.; et al. 2023. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 4015--4026

2023
[18]

Lin, T.-Y.; Maire, M.; Belongie, S.; Hays, J.; Perona, P.; Ramanan, D.; Doll \'a r, P.; and Zitnick, C. L. 2014. Microsoft coco: Common objects in context. In Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 1...

2014
[19]

Liu, Z.; Chen, H.; Feng, R.; Wu, S.; Ji, S.; Yang, B.; and Wang, X. 2021 a . Deep dual consecutive network for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 525--534

2021
[20]

Liu, Z.; Feng, R.; Chen, H.; Wu, S.; Gao, Y.; Gao, Y.; and Wang, X. 2022 a . Temporal feature alignment and mutual information maximization for video-based human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11006--11016

2022
[21]

Liu, Z.; Lin, Y.; Cao, Y.; Hu, H.; Wei, Y.; Zhang, Z.; Lin, S.; and Guo, B. 2021 b . Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, 10012--10022

2021
[22]

Liu, Z.; Mao, H.; Wu, C.-Y.; Feichtenhofer, C.; Darrell, T.; and Xie, S. 2022 b . A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11976--11986

2022
[23]

Liu, Z.; Wu, S.; Xu, C.; Wang, X.; Zhu, L.; Wu, S.; and Feng, F. 2022 c . Copy motion from one to another: Fake motion video generation. arXiv preprint arXiv:2205.01373

2022 arXiv
[24]

Newell, A.; Yang, K.; and Deng, J. 2016. Stacked hourglass networks for human pose estimation. In Computer Vision--ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, 483--499. Springer

2016
[25]

Pfister, T.; Charles, J.; and Zisserman, A. 2015. Flowing convnets for human pose estimation in videos. In Proceedings of the IEEE international conference on computer vision, 1913--1921

2015
[26]

Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Ommer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 10684--10695

2022
[27]

Sapp, B.; Toshev, A.; and Taskar, B. 2010. Cascaded models for articulated pose estimation. In Computer Vision--ECCV 2010: 11th European Conference on Computer Vision, Heraklion, Crete, Greece, September 5-11, 2010, Proceedings, Part II 11, 406--420. Springer

2010
[28]

Schmidtke, L.; Vlontzos, A.; Ellershaw, S.; Lukens, A.; Arichi, T.; and Kainz, B. 2021. Unsupervised human pose estimation through transforming shape templates. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2484--2494

2021
[29]

Shuai, C.; Zhong, J.; Wu, S.; Lin, F.; Wang, Z.; Ba, Z.; Liu, Z.; Cavallaro, L.; and Ren, K. 2023. Locate and verify: A two-stream network for improved deepfake detection. In Proceedings of the 31st ACM International Conference on Multimedia, 7131--7142

2023
[30]

Su, P.; Liu, Z.; Wu, S.; Zhu, L.; Yin, Y.; and Shen, X. 2021. Motion prediction via joint dependency modeling in phase space. In Proceedings of the 29th ACM international conference on multimedia, 713--721

2021
[31]

Sun, K.; Xiao, B.; Liu, D.; and Wang, J. 2019. Deep high-resolution representation learning for human pose estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 5693--5703

2019
[32]

Tian, X.; Zhang, Z.; Lin, S.; Qu, Y.; Xie, Y.; and Ma, L. 2021. Farewell to mutual information: Variational distillation for cross-modal person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1522--1531

2021
[33]

Tse, T. H. E.; Kim, K. I.; Leonardis, A.; and Chang, H. J. 2022. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 1664--1674

2022
[34]

N.; Kaiser, .; and Polosukhin, I

Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, .; and Polosukhin, I. 2017. Attention is all you need. Advances in neural information processing systems, 30

2017
[35]

Wang, M.; Tighe, J.; and Modolo, D. 2020. Combining detection and tracking for human pose estimation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 11088--11096

2020
[36]

Wang, Y.; Li, K.; Li, Y.; He, Y.; Huang, B.; Zhao, Z.; Zhang, H.; Xu, J.; Liu, Y.; Wang, Z.; et al. 2022. Internvideo: General video foundation models via generative and discriminative learning. arXiv preprint arXiv:2212.03191

2022 arXiv
[37]

Wei, S.-E.; Ramakrishna, V.; Kanade, T.; and Sheikh, Y. 2016. Convolutional pose machines. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 4724--4732

2016
[38]

Wu, S.; Chen, H.; Yin, Y.; Hu, S.; Feng, R.; Jiao, Y.; Yang, Z.; and Liu, Z. 2024 a . Joint-Motion Mutual Learning for Pose Estimation in Video. In Proceedings of the 32nd ACM International Conference on Multimedia, 8962--8971

2024
[39]

Wu, S.; Liu, Z.; Zhang, B.; Zimmermann, R.; Ba, Z.; Zhang, X.; and Ren, K. 2024 b . Do as I Do: Pose Guided Human Motion Copy. IEEE Transactions on Dependable and Secure Computing

2024
[40]

Xiao, B.; Wu, H.; and Wei, Y. 2018. Simple baselines for human pose estimation and tracking. In Proceedings of the European conference on computer vision (ECCV), 466--481

2018
[41]

Xiu, Y.; Li, J.; Wang, H.; Fang, Y.; and Lu, C. 2018. Pose Flow: Efficient online pose tracking. arXiv preprint arXiv:1802.00977

2018 arXiv
[42]

Xu, Y.; Zhang, J.; Zhang, Q.; and Tao, D. 2022. Vitpose: Simple vision transformer baselines for human pose estimation. Advances in Neural Information Processing Systems, 35: 38571--38584

2022
[43]

Yang, Y.; Chen, H.; Liu, Z.; Lyu, Y.; Zhang, B.; Wu, S.; Wang, Z.; and Ren, K. 2023. Action recognition with multi-stream motion modeling and mutual information maximization. arXiv preprint arXiv:2306.07576

2023 arXiv
[44]

Zhang, X.; Li, C.; Tong, X.; Hu, W.; Maybank, S.; and Zhang, Y. 2009. Efficient human pose estimation via parsing a tree structure based human model. In 2009 IEEE 12th International Conference on Computer Vision, 1349--1356. IEEE

2009
[45]

Zhu, X.; Hu, H.; Lin, S.; and Dai, J. 2019. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 9308--9316

2019
[46]

, " * write output.state after.block = add.period write newline

ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...
[47]

write newline

" write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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

[1] [1]

Andriluka, M.; Iqbal, U.; Insafutdinov, E.; Pishchulin, L.; Milan, A.; Gall, J.; and Schiele, B. 2018. Posetrack: A benchmark for human pose estimation and tracking. In Proceedings of the IEEE conference on computer vision and pattern recognition, 5167--5176

work page 2018

[2] [2]

Artacho, B.; and Savakis, A. 2020. Unipose: Unified human pose estimation in single images and videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7035--7044

work page 2020

[3] [3]

Bertasius, G.; Feichtenhofer, C.; Tran, D.; Shi, J.; and Torresani, L. 2019. Learning temporal pose estimation from sparsely-labeled videos. Advances in neural information processing systems, 32

work page 2019

[4] [4]

Chen, X.; Wang, X.; Zhou, J.; Qiao, Y.; and Dong, C. 2023. Activating more pixels in image super-resolution transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 22367--22377

2023

[5] [5]

Dai, J.; Qi, H.; Xiong, Y.; Li, Y.; Zhang, G.; Hu, H.; and Wei, Y. 2017. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, 764--773

2017

[6] [6]

Deng, J.; Dong, W.; Socher, R.; Li, L.-J.; Li, K.; and Fei-Fei, L. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, 248--255. Ieee

2009

[7] [7]

Doering, A.; Chen, D.; Zhang, S.; Schiele, B.; and Gall, J. 2022. Posetrack21: A dataset for person search, multi-object tracking and multi-person pose tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 20963--20972

work page 2022

[8] [8]

Dosovitskiy, A.; Beyer, L.; Kolesnikov, A.; Weissenborn, D.; Zhai, X.; Unterthiner, T.; Dehghani, M.; Minderer, M.; Heigold, G.; Gelly, S.; et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929

arXiv 2020