VISUALCENT: Visual Human Analysis using Dynamic Centroid Representation

Guanghui Wang; Niaz Ahmad; Youngmoon Lee

REVIEW 3 major objections 4 minor 46 references

VISUALCENT: Visual Human Analysis using Dynamic Centroid Representation

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read VISUALCENT combines keypoint detection and instance segmentation in one bottom-up network using dynamic keypoint centroids as grouping anchors.

desk verdict Reported bottom-up pose and segmentation numbers are strong, but the MaskCentroid grouping rule is never specified, so the central segmentation claim cannot be checked from the paper. read the letter →

arxiv 2504.19032 v1 pith:N6ORY7FC submitted 2025-04-26 cs.CV cs.AI

Niaz Ahmad , Youngmoon Lee , Guanghui Wang This is my paper · ORCID

classification cs.CVcs.AI

keywords humanposeestimationinstancesegmentationbottom-updynamiccentroidkeypointheatmapocclusionCOCOOCHuman

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

VISUALCENT is a unified bottom-up framework for multi-person human pose estimation and instance-level segmentation that does not need a person detector. The paper's central idea is to use a high-confidence predicted keypoint as a dynamic centroid (MaskCentroid) to which all mask pixels of that person are attached by embedding distance, so grouping is cheap and adjusts to movement and occlusion. The authors report that this design outperforms prior unified models such as PosePlusSeg and PersonLab on COCO and OCHuman benchmarks, and runs at a higher frame rate. The practical payoff is a scalable, real-time visual human analysis system for crowded or entangled scenes.

What carries the argument

The load-bearing mechanism is the dynamic centroid, used in two linked operations. KeyCentroid is a two-channel regression within the keypoint disk (radius $R=32$) that points toward the joint's location, refining the coarse heatmap response and providing high-confidence keypoint coordinates. MaskCentroid then selects one of those high-confidence keypoints per person as a learnable center of attraction for pixel clustering, grouping pixels by embedding distance to that center with the exponential affinity in equation (3). The dynamic adjustment of the centroid is what the paper credits for keeping segmentation stable under rapid body motion and severe occlusion, and it is also what keeps the runtime low, since each pixel only compares against a few anchors rather than all other pixels.

What would settle it

Re-run the static-versus-dynamic MaskCentroid comparison on the COCO validation set and on OCHuman with numeric mask AP, replacing the dynamic keypoint anchor with a fixed geometric centroid (the mean of the person's predicted keypoints or the static centroid used in PosePlusSeg). If the static variant matches or exceeds the dynamic variant's AP, then the dynamic anchor is not the source of the reported improvement.

Watch

Extended reading notes

Core claim

The core claim is that a single dynamic centroid per person, derived from a confidently detected keypoint, resolves the pixel-to-instance grouping problem in bottom-up segmentation. VISUALCENT first produces keypoint heatmaps with a disk representation (a radius-32 region around each joint) and a KeyCentroid regression that points from any pixel in the disk to the joint's exact location; these two outputs are combined to pick an optimal keypoint. One such keypoint per instance is then treated as a learnable MaskCentroid, and all mask pixels are clustered by embedding proximity using $\phi_j(e_i)=\exp(-\lVert e_i-\mu_j\rVert^2 / 2\sigma_j^2)$, where $\mu_j$ is the mean embedding of the instance. The paper claims this dynamic choice beats a fixed centroid when bodies move or occlude each other, and on COCO test-dev reaches 76.1 keypoint AP with ResNet-152 and 47.6 mask AP on the segmentation test set, along with gains on OCHuman.

Load-bearing premise

The method assumes that at least one keypoint per person is detected with high confidence even when bodies overlap or are occluded, because that single keypoint is the anchor that gathers all of the person's mask pixels; if no keypoint is reliable or the chosen one belongs to a different instance, the segmentation grouping has no correct center to cluster around.

Editorial extensions

If this is right

A single bottom-up model could replace the detect-then-segment cascade for human analysis, eliminating the extra compute and failure modes of a person detector.
The per-pixel grouping cost scales with the number of anchors, so crowded frames remain tractable without pairwise pixel comparisons.
Real-time video applications such as surveillance, human-robot interaction, and live sports analysis could run pose and mask estimation at interactive frame rates.
The dynamic-centroid mechanism appears to transfer across backbone capacities, since the paper shows gains with both ResNet-101 and ResNet-152.

Reading between the lines

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

The MaskCentroid idea is portable to other instance-level dense tasks, such as animal pose and part segmentation, whenever one reliable semantic keypoint per instance exists.
If the anchor is the single highest-confidence keypoint, segmentation accuracy becomes coupled to keypoint confidence calibration; a separate learned anchor-confidence could decouple them.
A quantitative ablation that varies the choice of anchor keypoint (e.g., highest confidence vs. a limb-specific joint) would clarify how much of the gain comes from the dynamic selection rule itself.
The method's limitation is likely a person with no visible keypoint at all; adding a fallback centroid, such as the mean of lower-confidence keypoints or a learned embedding center, would extend it to fully occluded instances.

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.

Desk Editor's Note

read the letter

Short version: the pose numbers are strong, the segmentation mechanism is not specified well enough to check, and the paper's headline claims would need a serious rewrite before they're believable.

What's actually new: KeyCentroid is a disk-constrained keypoint offset regression, a plausible variation on PersonLab's short-range offsets and CenterNet's center offsets. MaskCentroid is the more interesting idea: anchor mask clustering at a high-confidence keypoint so the anchor can move dynamically. The COCO keypoint results (76.1 AP with ResNet-152) and the OCHuman numbers are the kind of numbers that would move this subfield if they reproduce.

What the paper does well: it evaluates on public benchmarks, compares directly against a reasonable list of bottom-up baselines including its own PosePlusSeg, and the KHDR+-KeyCentroid ablation gives concrete numbers. That part is checkable and looks sincere.

Soft spots, in order. Biggest: the MaskCentroid clustering rule is never actually specified. Equations (2) and (3) define a mean pixel position and a Gaussian similarity to a mean embedding, but both formulas assume the instance N_j is already known. There is no inference-time assignment rule that takes a predicted high-confidence keypoint and produces a set of pixels for that instance, and no training loss is written for the mask centroid offset. Saying the keypoint is a learnable parameter is not an algorithm. So the central segmentation claim cannot be reproduced or even falsified from the manuscript. The stress-test note is right, and it's not a minor omission.

Next: numeric inconsistencies. The "10.0% improvement" over HGG is an absolute 10.0 AP point gap, which at the reported values is around 28% relative. The "3.5%" over MIPNet is likewise a point gap. Table III labels Pose+Seg as [1] in the table but [9] in the text, and reference [9] is Dantone et al., not PosePlusSeg. The reference list also duplicates [5]. These are likely carelessness, but they erode confidence in the reported deltas.

Third: the dynamic-vs-static MaskCentroid comparison is a figure with no numbers, and the FPS/real-time claim is a plot with no table. For a paper that leads with a real-time claim, that's under-evidenced.

Bottom line: the pose side is salvageable and worth a serious referee. The segmentation side, as written, is not verifiable. I would send it to review only with the expectation that the authors supply the assignment rule, the mask loss, and numeric ablations. It deserves referee time rather than desk rejection because the reported results, if real, are a clear within-subfield improvement and the system idea is coherent. For me personally, I'd cite the KeyCentroid pose result with caution; I wouldn't cite the MaskCentroid mechanism until it's written down.

Who it's for: bottom-up pose/segmentation researchers, especially those working on crowded real-time scenarios. They'd get something from the disk-offset idea; they'd need to wait for a revised version on the clustering.

Recommendation: conditional acceptance after major revision, with the missing algorithm and data supplied.

Referee Report

3 major / 4 minor

Summary. The paper introduces VISUALCENT, a bottom-up framework for unified multi-person 2D pose estimation and instance segmentation. The method first detects keypoints using a keypoint heatmap with a disk representation and a 'KeyCentroid' regressed offset field, then groups mask pixels to instances using a 'MaskCentroid' defined as a high-confidence keypoint that serves as a dynamic clustering anchor. Evaluations are reported on COCO keypoint test-dev, COCO segmentation test, and OCHuman, with tables showing improvements over prior bottom-up methods such as PosePlusSeg and PersonLab. The abstract also claims real-time performance. The central technical novelty is the dynamic MaskCentroid, which is claimed to outperform a static centroid under occlusion and rapid motion.

Significance. If the reported results are reproducible and the dynamic-centroid mechanism is fully specified, the paper would make a useful contribution: it is a detector-free unified pose-and-segmentation model that reports competitive mAP on COCO and particularly strong gains on OCHuman, a highly occluded benchmark. The external evaluation on standard benchmarks and the direct comparisons with recent bottom-up methods are strengths. However, the paper's central claim about the dynamic MaskCentroid is currently under-specified, and the key ablation supporting it contains no numbers. The real-time claim is also not quantified. These gaps prevent the reader from verifying or building on the method, so the current version does not meet the evidentiary bar for its headline claims.

major comments (3)

[Section III-D (Dynamic MaskCentroid)] The inference-time assignment rule for MaskCentroid is never defined. Equations (2) and (3) describe a centroid F' and a similarity phi_j(e_i) using the known set N_j of pixels belonging to an instance, but the paper does not specify how a predicted high-confidence keypoint anchors pixels, how the similarity threshold or assignment procedure works, how overlapping instances are separated, or what happens when the anchoring keypoint is occluded or incorrectly detected. The text 'defining the highly confident keypoint as a learnable parameter' is not an algorithm. The training loss is also not written: the paper only says 'the MaskCentroid offset loss' is minimized (Section III). Without this specification, the segmentation pipeline cannot be reproduced or even falsified from the manuscript, which is a load-bearing gap for the central claim that dynamic MaskCentroids improve mask AP.
[Section IV-A, Fig. 6] The key ablation comparing Static MaskCentroid (SMc) and Dynamic MaskCentroid (DMc) is presented only as a figure with no numerical values. The text states that 'the exceptional performance of the proposed DMc approach demonstrates its effectiveness' but provides no AP numbers, no standard deviation, and no evaluation protocol details for Fig. 6. Since the dynamic-centroid design is the paper's main novelty, this ablation must be reported with concrete numbers (e.g., a table with AP, AP_50, AP_75) for the claim to be verifiable.
[Abstract and Section IV (Computational Cost)] The abstract claims 'real time performance advantages' and 'execution frame rate per second' advantages, and Section IV states that Fig. 5 shows 'high FPS' relative to sister models. However, no FPS table or per-model frame-rate numbers are given anywhere in the manuscript. Fig. 5 is a scatter plot without a clear quantitative legend. A real-time claim requires at least a table reporting FPS, hardware, batch size, and input resolution for each compared model; the current evidence is insufficient to assess the central speed claim.

minor comments (4)

[Section IV (OCHuman results)] The text says VISUALCENT 'improves 10.0% compare to HGG [18]' using the OCHuman test set; since the improvement is from 36.0 to 46.0 AP, this is a 10.0 percentage-point gain, not a 10.0% relative improvement. Please clarify the terminology.
[Table III] Table III lists PersonLab four times, with two rows per backbone and identical backbone labels but different AP values. Please label which rows are single-scale and which are multi-scale, and reconcile the caption footnote with the table entries.
[References] In the Segmentation Results paragraph, 'Pose+Seg [9]' appears to be a citation error: reference [9] is Dantone et al., not PosePlusSeg; the PosePlusSeg reference is [1]. Similarly, in Section IV the MIPNet citation appears as [19] while the reference list shows MIPNet as [20].
[Throughout] The method name is spelled inconsistently: 'VISUALCENT' in the abstract and 'VISUAL CENT' in the body and title. Also, 'MaskCenroid' appears in Section I. Please unify the spelling.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is evaluated on external public benchmarks and the reported improvements are empirical comparisons, not derivations that fold the target result into the construction.

full rationale

The paper makes empirical claims about mAP and FPS that are supported by evaluations on COCO and OCHuman test sets, with the model trained end-to-end from public supervision. Equation (1) is a standard Gaussian-shaped regression target for keypoint displacements, Eq. (2) defines an instance centroid given its pixels, and Eq. (3) is a similarity kernel in embedding space; none of these equations equates a fitted parameter with a claimed prediction. The only overlapping-author citation, [1] PosePlusSeg, is used transparently as a baseline and as the contrast for the dynamic-centroid idea, but it is not load-bearing for the validity of the reported benchmark numbers, which rest on external test sets. The manuscript does leave the MaskCentroid inference-time assignment rule and the MaskCentroid offset loss unspecified, so the segmentation mechanism cannot be fully checked from the paper; that is a completeness and reproducibility gap, not circularity, because no reported result is forced by construction.

Assumptions & free parameters 2 free parameters · 4 assumptions · 2 invented entities

The central method rests on standard deep learning assumptions (supervised training on COCO, convolutional backbones) plus two domain assumptions: a fixed-radius keypoint disk captures keypoint localization, and a single predicted keypoint can anchor mask clustering for an entire person. The dynamic-centroid learning rule is asserted rather than derived. No new physical entity is proposed; the only invented constructs are the KeyCentroid and MaskCentroid representations, whose independent evidence is limited to the paper's own experiments.

free parameters (2)

Keypoint disk radius R = 32
Set by hand in Section III to normalize the KeyCentroid response; Eq. (1) uses it as denominator, and a different radius would change the target fields and likely the results.
sigma_j (embedding bandwidth in Eq. 3) = not specified
Controls the soft assignment width for MaskCentroid clustering; no value is given in the paper, so it is an unstated tuning choice.

assumptions (4)

domain assumption Keypoint disks of fixed radius R=32 can represent all keypoint regions across scales and occlusions.
Section III defines KHDR with a fixed disk; no scale-adaptive mechanism is demonstrated.
domain assumption A single high-confidence keypoint per instance is a sufficient cluster anchor for all mask pixels.
Section III Dynamic MaskCentroid, Eq. (3) assigns pixels by proximity to the instance centroid; if anchors are wrong, segmentation fails, and no fallback is described.
ad hoc to paper The dynamic centroid can be optimized by treating the high-confidence keypoint as a learnable parameter.
The learning rule is asserted in Section III but no derivation or loss term for the 'learnable parameter' is given; the paper's ablation (Fig. 6) has no numeric support.
domain assumption Training on COCO transfers to OCHuman without domain adaptation.
Evaluation in Section IV applies the COCO-trained model to OCHuman; no specific transfer mechanism is described.

invented entities (2)

KeyCentroid (Kc)
purpose: A two-channel per-keypoint vector field inside the keypoint disk that regresses each pixel toward the true keypoint coordinate, used to refine keypoint localization.
It is an algorithmic component whose benefit is shown only by the paper's own ablation in Table V; there is no external falsifiable prediction from the construct itself.
MaskCentroid (Mc)
purpose: A high-confidence predicted keypoint used as a dynamic anchor for clustering mask pixels into person instances in embedding space.
The central new component; its advantage over a static centroid is only asserted via Fig. 6, which contains no numerical values, so it has no independent evidence outside the paper's claims.

how reviews work

0 comments

Cite this review

Pith. "Pith review of VISUALCENT: Visual Human Analysis using Dynamic Centroid Representation." pith.science (2026). https://pith.science/paper/N6ORY7FC

@misc{pith2026250419032,
  author       = {Pith},
  title        = {Pith review of: VISUALCENT: Visual Human Analysis using Dynamic Centroid Representation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N6ORY7FC}},
  note         = {Machine review of arXiv:2504.19032}
}

read the original abstract

We introduce VISUALCENT, a unified human pose and instance segmentation framework to address generalizability and scalability limitations to multi person visual human analysis. VISUALCENT leverages centroid based bottom up keypoint detection paradigm and uses Keypoint Heatmap incorporating Disk Representation and KeyCentroid to identify the optimal keypoint coordinates. For the unified segmentation task, an explicit keypoint is defined as a dynamic centroid called MaskCentroid to swiftly cluster pixels to specific human instance during rapid changes in human body movement or significantly occluded environment. Experimental results on COCO and OCHuman datasets demonstrate VISUALCENTs accuracy and real time performance advantages, outperforming existing methods in mAP scores and execution frame rate per second. The implementation is available on the project page.

Figures

Figures reproduced from arXiv: 2504.19032 by the authors.

**Figure 1.** Visual performance of VISUALCENT. (a) represents KeyCentroid to identify precise keypoint coordinates. (b) represents MaskCentroid with dynamic high confident keypoint to cluster the mask pixels to the correct instance. (c) represents an illustration of a binary mask generated with the help of MaskCentroid (d) exemplifies the instance-level segmentation. (e) shows a unified representation of human pose and estimatio… view at source ↗

**Figure 2.** The overview of the proposed system. F ∗ indicates the KeyCentroid Kc operation. F ′ indicates the MaskCentroid Mc operation using the Explicit keypoint Ek predicted by Kc. several major differences that make VISUALCENT more effective, scalable, and real-time. First, they rely on static features to detect or group keypoints by using greedy decoding; in contrast, VISUALCENT introduces KeyCentroid that calculates the… view at source ↗

**Figure 4.** The segmentation map is created using MaskCentroid, a dynamic [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: Computational cost with the representative sister models. Models are tested on a single Titan RTX [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 40 canonical work pages

[1]

Ahmad, J

N. Ahmad, J. Khan, J. Y . Kim, and Y . Lee. Joint Human Pose Estimation and Instance Segmentation with PosePlusSeg. In AAAI, 2022

work page 2022
[9]

Dantone, J

M. Dantone, J. Gall, C. Leistner, and L. Van Gool. Human pose estimation using body parts dependent joint regressors. In CVPR, 2013

work page 2013
[5]

Z. Cao, T. Simon, S.-E. Wei, and Y . Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 7291–7299, 2017

work page 2017
[2]

Bolya, C

D. Bolya, C. Zhou, F. Xiao, and Y . J. Lee. Yolact: Real-time instance segmentation. In ICCV, 2019

work page 2019
[3]

Y . Cai, Z. Wang, Z. Luo, B. Yin, A. Du, H. Wang, X. Zhang, X. Zhou, E. Zhou, and J. Sun. Learning delicate local representations for multi- person pose estimation. In ECCV, 2020

work page 2020
[4]

Z. Cao, T. Simon, S.-E. Wei, and Y . Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017

work page 2017
[6]

Y . Chen, Z. Wang, Y . Peng, Z. Zhang, G. Yu, and J. Sun. Cascaded pyramid network for multi-person pose estimation. In CVPR, 2018

work page 2018
[7]

Cheng, B

B. Cheng, B. Xiao, J. Wang, H. Shi, T. S. Huang, and L. Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, 2020

work page 2020

Show all 46 references

[8]

J. Dai, K. He, Y . Li, S. Ren, and J. Sun. Instance-sensitive fully convolutional networks. In ECCV, 2016

2016
[10]

H.-S. Fang, S. Xie, Y .-W. Tai, and C. Lu. Rmpe: Regional multi-person pose estimation. In ICCV, 2017

2017
[11]

Gidaris, Tompson, and K

George, Zhu, Chen, J. Gidaris, Tompson, and K. Murphy. Personlab: Person pose estimation and instance segmentation with a bottom-up, part-based, geometric embedding model. In ECCV, 2018

2018
[12]

G. Han, C. Song, S. Wang, H. Wang, E. Chen, and G. Wang. Occluded human pose estimation based on limb joint augmentation. Neural Computing and Applications , 37(3):1241–1253, 2025

2025
[13]

He, Xiangyu, Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016

2016
[14]

K. He, G. Gkioxari, P. Doll ´ar, and R. Girshick. Mask r-cnn. In ICCV, 2017

2017
[15]

Huang, M

S. Huang, M. Gong, and D. Tao. A coarse-fine network for keypoint localization. In ICCV, 2017

2017
[16]

Insafutdinov, L

E. Insafutdinov, L. Pishchulin, B. Andres, M. Andriluka, and B. Schiele. Deepercut: A deeper, stronger, and faster multi-person pose estimation model. In ECCV, 2016

2016
[17]

Jiang, S

W. Jiang, S. Jin, W. Liu, C. Qian, P. Luo, and S. Liu. Posetrans: A simple yet effective pose transformation augmentation for human pose estimation. In European Conference on Computer Vision , pages 643–659. Springer, 2022

2022
[18]

S. Jin, W. Liu, E. Xie, W. Wang, C. Qian, W. Ouyang, and P. Luo. Differentiable hierarchical graph grouping for multi-person pose esti- mation. In ECCV. Springer, 2020

2020
[19]

Khirodkar, V

R. Khirodkar, V . Chari, A. Agrawal, and A. Tyagi. Multi-hypothesis pose networks: Rethinking top-down pose estimation. arXiv preprint arXiv:2101.11223, 2021

2021 arXiv
[20]

Khirodkar, V

R. Khirodkar, V . Chari, A. Agrawal, and A. Tyagi. Multi-instance pose networks: Rethinking top-down pose estimation. In ICCV, 2021

2021
[21]

Kocabas, S

M. Kocabas, S. Karagoz, and E. Akbas. Multiposenet: Fast multi- person pose estimation using pose residual network. In ECCV, 2018

2018
[22]

Kreiss, L

S. Kreiss, L. Bertoni, and A. Alahi. Pifpaf: Composite fields for human pose estimation. In CVPR, 2019

2019
[23]

W. Li, Z. Wang, B. Yin, Q. Peng, Y . Du, T. Xiao, G. Yu, H. Lu, Y . Wei, and J. Sun. Rethinking on multi-stage networks for human pose estimation. arXiv preprint arXiv:1901.00148 , 2019

1901 arXiv
[24]

T. Lin, M. Maire, S. Belongie, James, Perona, Deva, Piotr, and Lawrence. Microsoft coco: Common objects in context. In ECCV, 2014

2014
[25]

T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie. Feature pyramid networks for object detection. In CVPR, 2017

2017
[26]

S. Liu, L. Qi, H. Qin, J. Shi, and J. Jia. Path aggregation network for instance segmentation. In CVPR, 2018

2018
[27]

J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In CVPR, 2015

2015
[28]

Newell, Z

A. Newell, Z. Huang, and J. Deng. Associative embedding: End-to-end learning for joint detection and grouping. In NeurIPS, 2017

2017
[29]

Newell, K

A. Newell, K. Yang, and J. Deng. Stacked hourglass networks for human pose estimation. In ECCV, 2016

2016
[30]

X. Nie, J. Feng, J. Zhang, and S. Yan. Single-stage multi-person pose machines. In ICCV, 2019

2019
[31]

Pishchulin, E

L. Pishchulin, E. Insafutdinov, S. Tang, B. Andres, M. Andriluka, P. V . Gehler, and B. Schiele. Deepcut: Joint subset partition and labeling for multi person pose estimation. In CVPR, 2016

2016
[32]

H. Qu, Y . Cai, L. G. Foo, A. Kumar, and J. Liu. A characteristic function-based method for bottom-up human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13009–13018, 2023

2023
[33]

S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015

2015
[34]

D. Shi, X. Wei, L. Li, Y . Ren, and W. Tan. End-to-end multi- person pose estimation with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 11069–11078, 2022

2022
[35]

K. Su, D. Yu, Z. Xu, X. Geng, and C. Wang. Multi-person pose estimation with enhanced channel-wise and spatial information. In CVPR, 2019

2019
[36]

Z. Tian, H. Chen, and C. Shen. Directpose: Direct end-to-end multi- person pose estimation. arXiv preprint arXiv:1911.07451 , 2019

1911 arXiv
[37]

J. Wang, X. Long, Y . Gao, E. Ding, and S. Wen. Graph-pcnn: Two stage human pose estimation with graph pose refinement. In ECCV, 2020

2020
[38]

T. Wang, L. Jin, Z. Wang, X. Fan, Y . Cheng, Y . Teng, J. Xing, and J. Zhao. Decenternet: Bottom-up human pose estimation via decentralized pose representation. In Proceedings of the 31st ACM International Conference on Multimedia , pages 1798–1808, 2023

2023
[39]

S.-E. Wei, V . Ramakrishna, T. Kanade, and Y . Sheikh. Convolutional pose machines. In CVPR, 2016

2016
[40]

Y . Wu, T. Marks, A. Cherian, S. Chen, C. Feng, G. Wang, and A. Sullivan. Unsupervised joint 3d object model learning and 6d pose estimation for depth-based instance segmentation. In Proceed- ings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019

2019
[41]

B. Xiao, H. Wu, and Y . Wei. Simple baselines for human pose estimation and tracking. In ECCV, 2018

2018
[42]

Y . Xiao, D. Yu, X. J. Wang, L. Jin, G. Wang, and Q. Zhang. Learning quality-aware representation for multi-person pose regression. In Pro- ceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 2822–2830, 2022

2022
[43]

W. Xu, C. Long, Y . Nie, and G. Wang. Disentangled representation learning for controllable person image generation. IEEE Transactions on Multimedia, 26:6065–6077, 2024

2024
[44]

N. Xue, T. Wu, G.-S. Xia, and L. Zhang. Learning local-global contex- tual adaptation for multi-person pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages 13065–13074, 2022

2022
[45]

Zhang, Z

J. Zhang, Z. Zhu, J. Lu, J. Huang, G. Huang, and J. Zhou. Simple: Single-network with mimicking and point learning for bottom-up human pose estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 3342–3350, 2021

2021
[46]

Zhang, R

S.-H. Zhang, R. Li, X. Dong, P. Rosin, Z. Cai, X. Han, D. Yang, H. Huang, and S.-M. Hu. Pose2seg: Detection free human instance segmentation. In CVPR, 2019

2019

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

Ahmad, J

N. Ahmad, J. Khan, J. Y . Kim, and Y . Lee. Joint Human Pose Estimation and Instance Segmentation with PosePlusSeg. In AAAI, 2022

work page 2022

[2] [9]

Dantone, J

M. Dantone, J. Gall, C. Leistner, and L. Van Gool. Human pose estimation using body parts dependent joint regressors. In CVPR, 2013

work page 2013

[3] [5]

Z. Cao, T. Simon, S.-E. Wei, and Y . Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 7291–7299, 2017

work page 2017

[4] [2]

Bolya, C

D. Bolya, C. Zhou, F. Xiao, and Y . J. Lee. Yolact: Real-time instance segmentation. In ICCV, 2019

work page 2019

[5] [3]

Y . Cai, Z. Wang, Z. Luo, B. Yin, A. Du, H. Wang, X. Zhang, X. Zhou, E. Zhou, and J. Sun. Learning delicate local representations for multi- person pose estimation. In ECCV, 2020

work page 2020

[6] [4]

Z. Cao, T. Simon, S.-E. Wei, and Y . Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In CVPR, 2017

work page 2017

[7] [6]

Y . Chen, Z. Wang, Y . Peng, Z. Zhang, G. Yu, and J. Sun. Cascaded pyramid network for multi-person pose estimation. In CVPR, 2018

work page 2018

[8] [7]

Cheng, B

B. Cheng, B. Xiao, J. Wang, H. Shi, T. S. Huang, and L. Zhang. Higherhrnet: Scale-aware representation learning for bottom-up human pose estimation. In CVPR, 2020

work page 2020