Pith. sign in

REVIEW 5 major objections 4 minor 53 references

Keypoints as Dynamic Centroids for Unified Human Pose and Segmentation

T0 review · 5 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that using detected keypoints as dynamic cluster centers lets a single bottom-up network produce accurate human joint locations and instance-level body segmentation in real time, without a person detector.

desk verdict A plausible bottom-up pose+segmentation idea, but the MaskCentroid loss is an identity as written and the ablation doesn't support the key claims. read the letter →

arxiv 2505.12130 v1 pith:RAAL5UTN submitted 2025-05-17 cs.CV cs.AI

Niaz Ahmad , Jawad Khan , Kang G. Shin , Youngmoon Lee , Guanghui Wang This is my paper · ORCID
classification cs.CVcs.AI
keywords humanposeestimationinstancesegmentationbottom-updynamiccentroidKeyMaskcrowdscenesocclusion
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 tries to establish that one bottom-up network can locate a person's joints and cut out the person's silhouette at the same time, without first running a separate person detector. The proposed KDC model treats detected keypoints as dynamic centroids: KeyCentroid refines joint coordinates inside a local disk, then MaskCentroid uses the high-confidence joints as moving cluster centers that pull mask pixels to the correct person in an embedding space. On COCO it reports a keypoint mAP of 76.1 and a segmentation mAP of 47.6 with a ResNet-152 backbone, along with strong results on the crowded CrowdPose and heavily occluded OCHuman benchmarks, while running in real time. If the claim holds, unified pose-and-segmentation becomes practical for live and crowded scenes without a detector's cost.

What carries the argument

The load-bearing machinery is the centroid pair. KeyCentroid places a disk $D_R(q)=\{p:\|p-q\|\le R\}$ around each joint; every pixel inside regresses a vector toward the joint's centroid, and pointwise Gaussian smoothing turns these into the final keypoint coordinates and confidence scores. MaskCentroid works in a learned embedding space where each mask pixel stores an offset toward its person's centroid; high-confidence keypoints act as dynamic centroids $C_i$, and pixel-to-instance assignment uses the Gaussian $\phi_j(e_i)=\exp(-\|e_i-C_i\|^2/2\sigma_j^2)$, with the dynamic variant (Eq. 7) replacing the centroid by the mean embedding. The combination lets one network cluster pixels to instances without a person detector, while the cluster centers can shift as keypoints move.

What would settle it

Run KDC on video frames where a person's visible body parts are separated by occlusion or a barrier, and track whether the dynamic centroid follows a fast-moving joint. If one person's pixels split into two clusters, or the mask lags when a joint moves quickly, the single-centroid assumption fails; the paper reports no such dynamic or disjoint-occlusion test.

Watch

Extended reading notes

Core claim

The paper's central claim is that a common centroid mechanism solves both tasks better than prior joint models. PoseNet produces keypoint heatmaps in a disk representation and uses KeyCentroid to regress each pixel inside the disk toward the joint's center, which the ablations credit for improved localization and confidence scores. SegNet then takes the resulting high-confidence keypoints as MaskCentroids, dynamic centers of attraction in the embedding space, and assigns each mask pixel to the instance whose centroid it is nearest. The reported gains—COCO keypoint AP 76.1 and segmentation AP 47.6, CrowdPose 74.5, OCHuman 46.3/46.0 keypoint and 58.3/59.6 segmentation—are attributed specifically to these two mechanisms rather than to the backbone, as the ablation study in Table 7 and the static-versus-dynamic comparison in Section 5.2 argue.

Load-bearing premise

The method assumes every person can be captured as one coherent blob of pixels clustered around a single center, even when the person is heavily occluded or moving fast, and the paper never directly tests that assumption in motion or with split-up bodies.

Editorial extensions

If this is right

  • A bottom-up, detector-free model reports higher COCO keypoint and segmentation AP than the top-down Mask R-CNN baseline, suggesting a person detector is not necessary for accurate joint pose plus mask output.
  • Because keypoints double as clustering centers, the network produces joints and masks from the same features, simplifying the architecture compared to separate pose and segmentation branches.
  • The reported runtime places the unified model in real-time territory with fewer parameters than Mask R-CNN, PersonLab, and Pose+Seg (Figure 6), opening live-use applications.
  • On the heavily occluded OCHuman benchmark the reported keypoint mAP of 46.3 (val) and 46.0 (test) exceeds the listed bottom-up baselines, indicating the grouping survives high occlusion levels.

Reading between the lines

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

  • The centroid mechanism is task-agnostic in principle: any articulated object with annotated keypoints and masks (animals, hands, robots) could use the same keypoint-driven clustering.
  • The paper motivates 'rapid dynamic movement' but evaluates only on static images; a video experiment with fast articulation would directly test whether the dynamic centroid actually tracks motion better than a static one.
  • The single-centroid-per-person assumption (Eq. 6) could be relaxed to a mixture of centroids per instance, which may better handle a person split into disconnected visible body parts, a case not covered by the benchmarks.
  • The disk radius R is fixed (R=32) in all reported experiments; making R depend on image scale or keypoint type could improve small- and large-person variation beyond the sigma tuning shown in Section 5.3.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 4 minor

Summary. The paper proposes KDC, a bottom-up unified model for human pose estimation and instance-level segmentation. PoseNet detects keypoints via a disk-based heatmap representation and refines them with a KeyCentroid mechanism, while SegNet uses high-confidence keypoints as dynamic MaskCentroids to cluster pixels in an embedding space. Experiments on COCO, CrowdPose, and OCHuman report strong or state-of-the-art accuracy, with real-time speed and no person detector. The central claim is that the KeyCentroid and dynamic MaskCentroid mechanisms are responsible for the reported gains.

Significance. If the reported results hold, KDC would be a valuable real-time, bottom-up unified pose-and-segmentation model; the keypoint results on COCO and CrowdPose are competitive, and the runtime/parameter comparisons are useful. The paper provides a quantitative ablation for the keypoint components (Table 7). However, the evidence for the MaskCentroid contribution is only qualitative, and the training objective for the offset branch is written in a degenerate form. The significance of the segmentation claims therefore cannot be assessed from the manuscript as it stands.

major comments (5)
  1. [Section 3.3, Eq. (5)] The offset loss is identically zero as written: e_i is defined by e_i = m_i + v_i, so ||e_i - (m_i + v_i)||_1 is always 0. This means the offset branch receives no gradient and the actual training signal for the pixel embeddings is unspecified. The authors must state the real loss and clarify how e_i is generated by the network.
  2. [Section 5.2 and Table 7] The only quantitative ablation (Table 7) reports keypoint AP; the SMc-vs-DMc comparison is presented only as a qualitative figure (Fig. 8) with no segmentation AP. Consequently, the +3.1 mask AP over Pose+Seg in Table 4 cannot be attributed to the MaskCentroid mechanism. A same-backbone, same-training ablation reporting mask metrics is needed to support the central claim.
  3. [Table 6] Table 6 compares KDC's 2D pose AP50 to projected 3D pose models (CRMH, ROMP) on CrowdPose. This conflates different tasks and evaluation protocols; the large margin is not a meaningful comparison. The authors should remove this comparison or explicitly label it as a different-task illustration rather than a competitive benchmark.
  4. [Section 1 and Section 3.3] The 'dynamic movement' motivation is never measured: all benchmarks are static images, and the dynamic-centroid benefit under rapid motion is only inferred from single-frame results. Please either add temporal/video experiments or qualify the claim that dynamic centroids help during rapid body movements.
  5. [Section 3.3, Eqs. (6)-(7)] Equation (7) defines the dynamic centroid as the mean of assigned pixel embeddings, but assignment depends on distances to that same centroid. The paper does not analyze this self-referential clustering, nor does it test the single-centroid assumption for highly occluded or disconnected person instances, which is directly relevant to the OCHuman evaluation. Add an experiment or analysis that isolates this assumption.
minor comments (4)
  1. [Section 1] The contribution list contains a typo: 'developement' should be 'development'.
  2. [Section 4] The learning rate is written as '0.1×e-4', which is ambiguous; it should be written as 1e-5 or similar.
  3. [References] Several references are malformed, e.g., [George et al., 2018] and [Golda and others, 2019]; the author lists and citation keys should be corrected.
  4. [Figures 8-10] Figures 8-10 are screenshots without clear axis labels or quantitative legends; they should be redrawn as standard plots so the ablation results are legible.

Circularity Check

1 steps flagged · score 6.0 of 10

The MaskCentroid offset loss (Eq. 5) is an identity, so the paper's main segmentation novelty is self-definitional; keypoint components and benchmark comparisons are otherwise not circular.

  1. self definitional [Section 3.3, Eq. (5)]
    "This is achieved by defining a pixel offset vector vi for each known pixel mi, so that the resulting embedding ei = mi + vi points from its respective instance centroid. We penalize pixel offset loss using the L1 loss function during model training: Loffset = 1/N Σ_{i=1}^N ∥ei − (mi + vi)∥1. (5)"

    The loss target (mi + vi) is exactly the definition of ei given two sentences earlier. Therefore Loffset = (1/N)Σ ||(mi + vi) − (mi + vi)||_1 = 0 identically, independent of the network parameters. The offset/embedding branch receives no gradient from this loss, so the paper's description of MaskCentroid learning is vacuous: the 'prediction' ei equals its own input by construction. Any segmentation accuracy attributed to Dynamic MaskCentroid therefore cannot be traced to the stated training objective; the claimed dynamic-centroid mechanism reduces to a definitional identity rather than a learned transformation.

full rationale

The keypoint side of the paper is self-contained: KeyCentroid is trained with a supervised L1 loss against ground-truth centroid vectors (Eq. 4), and the pose improvements are supported by a same-backbone ablation (Table 7). The SOTA comparisons on COCO, CrowdPose, and OCHuman are empirical and do not reduce to the paper's inputs. Self-citations to Pose+Seg and VisualCent are used as baselines and context, not as load-bearing authority. However, the MaskCentroid training objective in Eq. 5 is an identity: the text defines ei = mi + vi and then minimizes ||ei − (mi + vi)||_1, which is zero for any network output. No gradient for the offset branch follows, so the central dynamic-centroid mechanism has no stated learning signal from this loss. Since MaskCentroid is the paper's main novelty for instance segmentation and the claimed gains over Pose+Seg are attributed to it, the derivation of this component reduces by construction to a tautology. The dynamic centroid formula in Eq. 7 is a standard self-referential clustering update, but without a valid embedding loss it cannot support the reported segmentation gains. This is partial circularity in one central training objective, while the keypoint derivation remains independent.

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

The central claim rests on several hand-tuned parameters (disk radius, sigma ranges, assignment threshold, and unspecified loss weights) and on the unimodal centroid assumption for pixel clustering. No new physical entities are introduced. The losses themselves are standard supervised objectives, so the main empirical weight rests on the reported benchmark numbers, which are not independently verifiable from the paper.

free parameters (5)
  • keypoint disk radius R = 32
    Set for all keypoint heatmaps and KeyCentroids; no sensitivity analysis across datasets, though Fig. 7 shows confidence scores for R=8,16,32 on one example.
  • point-wise Gaussian sigma = 0.1-0.5 for HVK, 0.5-1.0 for LVK
    Chosen per keypoint class to improve mAP; treated as a design choice and tuned via ablation in Fig. 9.
  • instance-wise Gaussian sigma = 0.1 favored, range 0.1-0.5
    Selected based on ablation in Fig. 10 for fine masks in crowded scenes.
  • pixel assignment threshold = 0.5
    Pixels with probability exceeding 0.5 are assigned to instances in Sec. 3.4; no analysis of this threshold is provided.
  • embedding loss weights = not reported
    The losses L_heatmap, L_KeyCentroid, and L_offset are combined, but the weighting of each term is never specified.
assumptions (5)
  • standard math ResNet-101/152 pretrained on ImageNet is used as the backbone.
    The paper says ResNet-101 and ResNet-152 are used for training and testing, implicitly relying on standard pretrained initialization.
  • domain assumption COCO keypoint and segmentation labels are ground truth and sufficient to train a joint model.
    The model is trained end-to-end on COCO keypoint and segmentation training sets, assuming these annotations are consistent and complete enough for joint learning.
  • domain assumption A single centroid per human instance is sufficient to represent pixel embeddings for clustering.
    Eq. 6 and Eq. 7 assign pixels by distance to one centroid per instance; this unimodal assumption is not tested for occluded or articulated bodies.
  • ad hoc to paper High-confidence keypoints remain reliable anchors for pixel grouping under occlusion.
    The introduction claims robustness under significant occlusion, but no experiment varies occlusion level or verifies that keypoint confidence remains predictive in occluded cases.
  • ad hoc to paper Gaussian smoothing with hand-set sigma ranges preserves the information needed for keypoint and mask accuracy.
    The point-wise and instance-wise Gaussian optimizations in Sec. 3.1 and 3.3 introduce sigma ranges justified only by ablation on specific benchmarks, not by a principled derivation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Keypoints as Dynamic Centroids for Unified Human Pose and Segmentation." pith.science (2026). https://pith.science/paper/RAAL5UTN

@misc{pith2026250512130,
  author       = {Pith},
  title        = {Pith review of: Keypoints as Dynamic Centroids for Unified Human Pose and Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RAAL5UTN}},
  note         = {Machine review of arXiv:2505.12130}
}
read the original abstract

The dynamic movement of the human body presents a fundamental challenge for human pose estimation and body segmentation. State-of-the-art approaches primarily rely on combining keypoint heatmaps with segmentation masks but often struggle in scenarios involving overlapping joints or rapidly changing poses during instance-level segmentation. To address these limitations, we propose Keypoints as Dynamic Centroid (KDC), a new centroid-based representation for unified human pose estimation and instance-level segmentation. KDC adopts a bottom-up paradigm to generate keypoint heatmaps for both easily distinguishable and complex keypoints and improves keypoint detection and confidence scores by introducing KeyCentroids using a keypoint disk. It leverages high-confidence keypoints as dynamic centroids in the embedding space to generate MaskCentroids, allowing for swift clustering of pixels to specific human instances during rapid body movements in live environments. Our experimental evaluations on the CrowdPose, OCHuman, and COCO benchmarks demonstrate KDC's effectiveness and generalizability in challenging scenarios in terms of both accuracy and runtime performance. The implementation is available at: https://sites.google.com/view/niazahmad/projects/kdc.

Figures

Figures reproduced from arXiv: 2505.12130 by the authors.

Figure 1
Figure 1. PoseNet operation begins by generating keypoint [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overview of the proposed KDC model. PoseNet generates keypoint heatmaps and refines them with KeyCentroid [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. (a) presents Keypoint heatmap using keypoint disk, (b) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 8
Figure 8. Figure 8: The exceptional performance of the proposed [PITH_FULL_IMAGE:figures/full_fig_p006_8.png]
Figure 7
Figure 7. Figure 7: Left(L) and right(R) keypoint confidence score with varied [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 5
Figure 5. Figure 5: Visual results from various components of the system re [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 10
Figure 10. Figure 10: Instance-wise Gaus￾sian optimization with different σ. Small σ provides precise in￾stance mask. 6 Conclusion This paper considers the challenge of unified human pose estimation and instance-level segmentation, particularly in complex multi-person dynamic movement scen…
Figure 6
Figure 6. Figure 6: Computational cost with the representative sister models. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

53 extracted references · 50 canonical work pages

  1. [1]

    Joint Human Pose Estimation and Instance Segmentation with PosePlusSeg

    [Ahmad et al., 2022] Niaz Ahmad, Jawad Khan, Jeremy Yuhyun Kim, and Youngmoon Lee. Joint Human Pose Estimation and Instance Segmentation with PosePlusSeg. In AAAI,

    work page 2022

  2. [9]

    Gaussian kernel smoothing

    [Chung, 2020] Moo K Chung. Gaussian kernel smoothing. arXiv preprint arXiv:2007.09539,

    arXiv 2020

  3. [10]

    Instance-sensitive fully convolutional networks

    [Dai et al., 2016] Jifeng Dai, Kaiming He, Yi Li, Shaoqing Ren, and Jian Sun. Instance-sensitive fully convolutional networks. In ECCV,

    work page 2016

  4. [13]

    Bottom-up human pose estimation via disentangled keypoint reg

    [Geng et al., 2021] Zigang Geng, Ke Sun, Bin Xiao, Zhaox- iang Zhang, and Jingdong Wang. Bottom-up human pose estimation via disentangled keypoint reg. In CVPR,

    work page 2021

  5. [14]

    Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model

    [George et al., 2018] George, Zhu, Chen, Jonathan Gidaris, Tompson, and Kevin Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In ECCV,

    work page 2018

  6. [15]

    Human pose estima- tion for real-world crowded scenarios

    [Golda and others, 2019] Golda et al. Human pose estima- tion for real-world crowded scenarios. In A VSS,

    work page 2019

  7. [16]

    Pi-net: Pose interacting network for multi-person monocular 3d pose estimation

    [Guo et al., 2021] Wen Guo, Enric Corona, Francesc Moreno-Noguer, and Xavier Alameda-Pineda. Pi-net: Pose interacting network for multi-person monocular 3d pose estimation. In WACV,

    work page 2021

  8. [17]

    Occluded human pose estimation based on limb joint augmentation

    [Han et al., 2025] Gangtao Han, Chunxiao Song, Song Wang, Hao Wang, Enqing Chen, and Guanghui Wang. Occluded human pose estimation based on limb joint augmentation. Neural Computing and Applications , 37(3):1241–1253,

    work page 2025

Show all 53 references
  1. [18]

    Deep residual learning for image recognition

    [He et al., 2016] He, Xiangyu, Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR,

    2016

  2. [19]

    Mask r-cnn

    [He et al., 2017] Kaiming He, Georgia Gkioxari, Piotr Doll´ar, and Ross Girshick. Mask r-cnn. In ICCV,

    2017

  3. [20]

    A coarse-fine network for keypoint local- ization

    [Huang et al., 2017] Shaoli Huang, Mingming Gong, and Dacheng Tao. A coarse-fine network for keypoint local- ization. In ICCV,

    2017

  4. [21]

    Deepercut: A deeper, stronger, and faster multi-person pose estimation model

    [Insafutdinov et al., 2016] Eldar Insafutdinov, Leonid Pishchulin, Bjoern Andres, Mykhaylo Andriluka, and Bernt Schiele. Deepercut: A deeper, stronger, and faster multi-person pose estimation model. In ECCV,

    2016

  5. [22]

    Posetrans: A simple yet effective pose transformation augmentation for human pose estimation

    [Jiang et al., 2022] Wentao Jiang, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, and Si Liu. Posetrans: A simple yet effective pose transformation augmentation for human pose estimation. In ECCV,

    2022

  6. [23]

    Differ- entiable hierarchical graph grouping for multi-person pose estimation

    [Jin et al., 2020] Sheng Jin, Wentao Liu, Enze Xie, Wenhai Wang, Chen Qian, Wanli Ouyang, and Ping Luo. Differ- entiable hierarchical graph grouping for multi-person pose estimation. In ECCV,

    2020

  7. [24]

    Multi instance pose nets: Rethinking topdown pose estimation

    [Khirodkar et al., 2021] Rawal Khirodkar, Visesh Chari, Amit Agrawal, and Ambrish Tyagi. Multi instance pose nets: Rethinking topdown pose estimation. InICCV,

    2021

  8. [25]

    Multiposenet: Fast multi-person pose estimation using pose residual network

    [Kocabas et al., 2018] Muhammed Kocabas, Salih Karagoz, and Emre Akbas. Multiposenet: Fast multi-person pose estimation using pose residual network. In ECCV,

    2018

  9. [26]

    Pifpaf: Composite fields for human pose estimation

    [Kreiss et al., 2019] Sven Kreiss, Lorenzo Bertoni, and Alexandre Alahi. Pifpaf: Composite fields for human pose estimation. In CVPR,

    2019

  10. [27]

    Crowdpose: Efficient crowded scenes pose estimation and a new benchmark

    [Li et al., 2019] Jiefeng Li, Can Wang, Hao Zhu, Yihuan Mao, Hao-Shu Fang, and Cewu Lu. Crowdpose: Efficient crowded scenes pose estimation and a new benchmark. In CVPR,

    2019

  11. [28]

    Mi- crosoft coco: Common objects in context

    [Lin et al., 2014] Tsung Lin, Michael Maire, Serge Be- longie, James, Perona, Deva, Piotr, and Lawrence. Mi- crosoft coco: Common objects in context. In ECCV,

    2014

  12. [30]

    Path aggregation network for instance seg- mentation

    [Liu et al., 2018] Shu Liu, Lu Qi, Haifang Qin, Jianping Shi, and Jiaya Jia. Path aggregation network for instance seg- mentation. In CVPR,

    2018

  13. [31]

    Group pose: A simple baseline for end-to-end multi-person pose estima- tion

    [Liu et al., 2023] Huan Liu, Qiang Chen, Zichang Tan, Jiang-Jiang Liu, Jian Wang, Xiangbo Su, Xiaolong Li, Kun Yao, Junyu Han, Errui Ding, et al. Group pose: A simple baseline for end-to-end multi-person pose estima- tion. In ICCV,

    2023

  14. [35]

    Fcpose: Fully convolutional multi-person pose estimation with dynamic instance-aware convolu- tions

    [Mao et al., 2021] Weian Mao, Zhi Tian, Xinlong Wang, and Chunhua Shen. Fcpose: Fully convolutional multi-person pose estimation with dynamic instance-aware convolu- tions. In CVPR,

    2021

  15. [36]

    Single-stage multi-person pose machines

    [Nie et al., 2019] Xuecheng Nie, Jiashi Feng, Jianfeng Zhang, and Shuicheng Yan. Single-stage multi-person pose machines. In ICCV,

    2019

  16. [37]

    Towards accurate multi- person pose estimation in the wild

    [Papandreou et al., 2017] George Papandreou, Tyler Zhu, Nori Kanazawa, Alexander Toshev, Jonathan Tompson, Chris Bregler, and Kevin Murphy. Towards accurate multi- person pose estimation in the wild. In CVPR,

    2017

  17. [38]

    Deepcut: Joint subset partition and labeling for multi person pose estimation

    [Pishchulin et al., 2016] Leonid Pishchulin, Eldar Insafutdi- nov, Siyu Tang, Bjoern Andres, Mykhaylo Andriluka, Pe- ter V Gehler, and Bernt Schiele. Deepcut: Joint subset partition and labeling for multi person pose estimation. In CVPR,

    2016

  18. [39]

    [Qu et al., 2023] Haoxuan Qu, Yujun Cai, Lin Geng Foo, Ajay Kumar, and Jun Liu. A char. function based method for bottom-up human pose estimation. In CVPR,

    2023

  19. [40]

    Faster r-cnn: Towards real-time object detection with region proposal nets

    [Ren et al., 2015] Shaoqing Ren, Kaiming He, Ross Gir- shick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal nets. In NeurIPS,

    2015

  20. [41]

    End-to-end multi-person pose estima- tion with transformers

    [Shi et al., 2022] Dahu Shi, Xing Wei, Liangqi Li, Ye Ren, and Wenming Tan. End-to-end multi-person pose estima- tion with transformers. In CVPR,

    2022

  21. [42]

    Multi-person pose estimation with enhanced channel-wise and spatial information

    [Su et al., 2019] Kai Su, Dongdong Yu, Zhenqi Xu, Xin Geng, and Changhu Wang. Multi-person pose estimation with enhanced channel-wise and spatial information. In CVPR,

    2019

  22. [43]

    Integral human pose regression

    [Sun et al., 2018] Xiao Sun, Bin Xiao, Fangyin Wei, Shuang Liang, and Yichen Wei. Integral human pose regression. In ECCV,

    2018

  23. [44]

    Directpose: Direct end-to-end multi-person pose estima- tion

    [Tian et al., 2019] Zhi Tian, Hao Chen, and Chunhua Shen. Directpose: Direct end-to-end multi-person pose estima- tion. arXiv preprint arXiv:1911.07451,

    2019 arXiv

  24. [45]

    Contextual instance decoupling for robust multi-person pose estimation

    [Wang and Zhang, 2022] Dongkai Wang and Shiliang Zhang. Contextual instance decoupling for robust multi-person pose estimation. In CVPR,

    2022

  25. [46]

    Decenternet: Bottom-up human pose estimation via decentralized pose representation

    [Wang et al., 2023] Tao Wang, Lei Jin, Zhang Wang, Xiao- jin Fan, Yu Cheng, Yinglei Teng, Junliang Xing, and Jian Zhao. Decenternet: Bottom-up human pose estimation via decentralized pose representation. In MM,

    2023

  26. [47]

    Convolutional pose machines

    [Wei et al., 2016] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR,

    2016

  27. [48]

    Simple baselines for human pose estimation and tracking

    [Xiao et al., 2018] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV,

    2018

  28. [49]

    Learning local-global contextual adapta- tion for multi-person pose estimation

    [Xue et al., 2022] Nan Xue, Tianfu Wu, Gui-Song Xia, and Liangpei Zhang. Learning local-global contextual adapta- tion for multi-person pose estimation. In CVPR,

    2022

  29. [50]

    Explicit box detection unifies end-to-end multi-person pose estimation

    [Yang et al., 2023] Jie Yang, Ailing Zeng, Shilong Liu, Feng Li, Ruimao Zhang, and Lei Zhang. Explicit box detection unifies end-to-end multi-person pose estimation. arXiv preprint arXiv:2302.01593,

    2023 arXiv

  30. [51]

    Pose2seg: Detection free human instance segmentation

    [Zhang et al., 2019] Song-Hai Zhang, Ruilong Li, Xin Dong, Paul Rosin, Zixi Cai, Xi Han, Dingcheng Yang, Haozhi Huang, and Shi-Min Hu. Pose2seg: Detection free human instance segmentation. In CVPR,

    2019

  31. [52]

    Simple: Single- network with mimicking and point learning for bottom-up human pose estimation

    [Zhang et al., 2021] Jiabin Zhang, Zheng Zhu, Jiwen Lu, Junjie Huang, Guan Huang, and Jie Zhou. Simple: Single- network with mimicking and point learning for bottom-up human pose estimation. In AAAI,

    2021

  32. [53]

    Rethinking pose es- timation in crowds: overcoming the detection information bottleneck and ambiguity

    [Zhou et al., 2023] Mu Zhou, Lucas Stoffl, Mackenzie Wey- gandt Mathis, and Alexander Mathis. Rethinking pose es- timation in crowds: overcoming the detection information bottleneck and ambiguity. In ICCV, 2023

    2023

  33. [2013]

    Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time

    [Fang et al., ] Hao-Shu Fang, Jiefeng Li, Hongyang Tang, Chao Xu, Haoyi Zhu, Yuliang Xiu, Yong-Lu Li, and Cewu Lu. Alphapose: Whole-body regional multi-person pose estimation and tracking in real-time. TPAMI. [Fang et al., 2017] Hao-Shu Fang, Shuqin Xie, Yu-Wing Tai, and Cewu ...

    2017

  34. [2014]

    Fea- ture pyramid nets for object detection

    [Lin et al., 2017] Tsung-Yi Lin, Piotr Doll´ar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Fea- ture pyramid nets for object detection. In CVPR,

    2017

  35. [2015]

    Rtmo: Towards high- performance one-stage real-time multi-person pose esti- mation

    [Lu et al., 2024] Peng Lu, Tao Jiang, Yining Li, Xiangtai Li, Kai Chen, and Wenming Yang. Rtmo: Towards high- performance one-stage real-time multi-person pose esti- mation. In CVPR,

    2024

  36. [2016]

    Human pose estimation using body parts dependent joint regressors

    [Dantone et al., 2013] Matthias Dantone, Juergen Gall, Christian Leistner, and Luc Van Gool. Human pose estimation using body parts dependent joint regressors. In CVPR,

    2013

  37. [2017]

    Cas- caded pyramid network for multi-person pose estimation

    [Chen et al., 2018] Yilun Chen, Zhicheng Wang, Yuxiang Peng, Zhiqiang Zhang, Gang Yu, and Jian Sun. Cas- caded pyramid network for multi-person pose estimation. In CVPR,

    2018

  38. [2018]

    Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation

    [Cheng et al., 2020] Bowen Cheng, Bin Xiao, Jingdong Wang, Honghui Shi, Thomas S Huang, and Lei Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR,

    2020

  39. [2019]

    The center of attention: Center-keypoint grouping via attention for multi-person pose estimation

    [Bras´o et al., 2021] Guillem Bras´o, Nikita Kister, and Laura Leal-Taix´e. The center of attention: Center-keypoint grouping via attention for multi-person pose estimation. In ICCV,

    2021

  40. [2020]

    Realtime multi-person 2d pose estimation using part affinity fields

    [Cao et al., 2017] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR,

    2017

  41. [2021]

    Learning delicate local representations for multi-person pose esti- mation

    [Cai et al., 2020] Yuanhao Cai, Zhicheng Wang, Zhengx- iong Luo, Binyi Yin, Angang Du, Haoqian Wang, Xiangyu Zhang, Xinyu Zhou, Erjin Zhou, and Jian Sun. Learning delicate local representations for multi-person pose esti- mation. In ECCV,

    2020

  42. [2022]

    Visualcent: Visual human analysis us- ing dynamic centroid representation

    [Ahmad et al., 2025] Niaz Ahmad, Youngmoon Lee, and Guanghui Wang. Visualcent: Visual human analysis us- ing dynamic centroid representation. In FG. IEEE,

    2025

  43. [2023]

    Fully convolutional networks for seman- tic segmentation

    [Long et al., 2015] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for seman- tic segmentation. In CVPR,

    2015

  44. [2024]

    Rethink- ing the heatmap regression for bottom-up human pose es- timation

    [Luo et al., 2021] Zhengxiong Luo, Zhicheng Wang, Yan Huang, Liang Wang, Tieniu Tan, and Erjin Zhou. Rethink- ing the heatmap regression for bottom-up human pose es- timation. In CVPR,

    2021

  45. [2025]

    Yolact: Real-time instance segmenta- tion

    [Bolya et al., 2019] Daniel Bolya, Chong Zhou, Fanyi Xiao, and Yong Jae Lee. Yolact: Real-time instance segmenta- tion. In ICCV,

    2019

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.