Pith. sign in

REVIEW 5 major objections 6 minor 37 references

Part Segmentation of Human Meshes via Multi-View Human Parsing

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Backprojecting 2D parsers yields 3D part segmentation

desk verdict Useful pipeline, but the accuracy claims are circular: the evaluation uses the same pseudo-labels that trained the model, so the reported mIoU is not established. read the letter →

arxiv 2507.18655 v4 pith:MBL4APX4 submitted 2025-07-22 cs.CV eess.IV

James Dickens , Kamyar Hamad This is my paper · ORCID
classification cs.CVeess.IV
keywords 3Dhumanmeshsegmentationparsingpointcloudsemanticmulti-viewbackprojectionpseudo-ground-truthlabelingfarthestsamplingTransformerorientationalignment
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

This paper aims to show that per-vertex semantic segmentation of dense, real-world human meshes can be obtained without hand-labeled 3D data and without textures. The authors build a pseudo-ground-truth pipeline: each mesh is aligned to a canonical upright, front-facing pose, rendered from several viewpoints, and labeled by off-the-shelf 2D human parsers, after which the 2D label maps are backprojected to vertices and cleaned by clustering. A windowed farthest-point sampling scheme downsamples meshes to 10,000 or 100,000 points, and a geometry-only PointTransformer learns to predict part labels, with nearest-neighbor upsampling back to the original vertex count. On THuman2.1, the best configuration reaches 74.4 mean IoU with Sapiens v1 labels at 100,000 points.

What carries the argument

The load-bearing object is the pseudo-label production loop: keypoint-based canonical alignment places the mesh in a known orientation; a triangle-ID shader lets each rendered pixel vote for a specific vertex; multiple views are aggregated by per-label voting; and DBSCAN followed by k-NN propagation removes noisy or disconnected label islands. Around this loop sits a sampling scheme that Morton-serializes normalized point coordinates into windows and applies iterative farthest point sampling window-by-window, cutting runtime from O(kN) to O(kN/W), then feeds a PointTransformer self-attention network on geometry only, with nearest-neighbor upsampling to full mesh resolution.

What would settle it

Take a sample of THuman2.1 scans, have human annotators label vertices for the same label sets, and compute the model's mIoU against those human labels; if the score on small parts such as hands, feet, lips, or teeth is far below the reported numbers, the pseudo-ground-truth shortcut, rather than geometry learning, is carrying the result. A cheaper check is to render scans in unusual or heavily occluded poses and see whether keypoint alignment and multi-view voting still produce spatially coherent labels.

Watch

Extended reading notes

Core claim

The paper's central claim is that multi-view 2D human parsing can serve as a reliable teacher for 3D geometry-only segmentation. With a keypoint-based pose-correction step, a triangle-ID rendering trick to associate pixels with vertices, and DBSCAN plus k-NN denoising, the backprojected labels from two 2D parsers are consistent enough to train a PointTransformer that reaches 74.4 mIoU and 96.8 accuracy on held-out THuman2.1 scans. The authors also claim that their Morton-sorted, windowed iterative farthest point sampling, with part-specific oversampling, is memory-efficient and preserves semantic structure for learning.

Load-bearing premise

The entire training and evaluation loop assumes the 2D human parsers' backprojected labels on rendered scans are correct enough to serve as ground truth, and the paper never checks them against human-verified 3D annotations, so systematic parser errors would inflate every reported number.

Editorial extensions

If this is right

  • Untextured meshes or models with missing texture maps can be semantically segmented using raw geometry alone.
  • Large scan collections become a source of automatic 3D part labels, removing the need for manual 3D annotation.
  • Increasing sampling density from 10,000 to 100,000 points consistently improves results in the best configuration, pointing to capacity for even finer supervision.
  • The windowed sampling strategy makes training on meshes with hundreds of thousands to millions of vertices practical on standard GPUs.
  • Because the pipeline does not require fitting a parametric body model, it can apply to diverse real-world scan shapes and clothing.

Reading between the lines

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

  • If the 2D parsers make systematic errors on rendered scans, the reported mIoU is optimistic; a small human-verified 3D label subset could calibrate or correct these numbers, a step the paper does not take.
  • The method's ceiling is set by the 2D parser's label space, so richer or open-vocabulary 2D segmenters could widen the 3D label set without changing the pipeline.
  • Nearest-neighbor upsampling likely blurs boundaries on small parts; a learned or surface-aware upsampler is a natural next step.
  • The approach should be stress-tested on non-canonical poses and heavy self-occlusion, because the keypoint alignment and multi-view voting assume a reasonably upright, mostly viewable subject.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The manuscript presents a pipeline for per-vertex semantic segmentation of THuman2.1 human meshes without texture information. It aligns meshes using PCA plus iterative YOLOv8-based pose correction, renders 12 views, segments them with M2FP or Sapiens, backprojects labels via a triangle-ID buffer, denoises with DBSCAN and k-NN, and downsamples using Morton-serialized windowed FPS before training a PointTransformer. Experiments report mIoU up to 74.4 on the Sapiens v1 label set after nearest-neighbor upsampling. The authors also release code and pre-processed data.

Significance. The pipeline addresses a practically useful gap—automatic 3D part labeling for high-resolution real scans—and the windowed FPS sampling idea is sensible for memory-limited training. The release of code and pre-processed data is a strength. However, the central accuracy claim is not currently supported because the pseudo-GT used for training is also used for evaluation, with no external validation, no baselines, and no error bars. If the evaluation is strengthened, the approach would be a useful contribution to geometry-based human parsing.

major comments (5)
  1. [Section 4.3, Table 4.2] The evaluation is circular. The training pseudo-labels and the test 'ground truth' are both produced by the same 2D parsing models (M2FP and Sapiens) via the Section 3.2 backprojection pipeline, so the reported mIoU measures how well the PointTransformer reproduces M2FP/Sapiens outputs, not the correctness of the 3D part segmentation. To support the abstract's accuracy claim, the authors must validate the pseudo-labels against an independent source—e.g., manually annotated per-vertex labels on a subset of THuman2.1, SMPL-X/SMPL part registrations, or the Human3D benchmark—and report per-class agreement. Without this, all numbers in Table 4.2 are upper bounds on agreement with the parsers.
  2. [Section 3.3, Eq. (4)] The stated formula for the number of windows, (N + 5000 − 1) (mod 5000), is incorrect; for example, N=10,000 gives 4999, not 2. The correct expression is ceil(N/5000). Also, Eq. (5) contains a self-referential neighborhood N (xj) and mixes xj and xi; the PointTransformer update should sum over neighbors of xi. These errors make the sampling description and the model formula unusable as written.
  3. [Section 4.3, Table 4.2] No baselines are reported. The central claim that the proposed windowed FPS sampling and PointTransformer pipeline is 'effective and accurate' requires comparison with at least random sampling, standard FPS (or PointNet++ style), and existing 3D human parsing methods (e.g., Human3D, open-vocabulary part segmentation). Without such comparisons the reported absolute mIoU values have no context.
  4. [Sections 4.2 and 4.3] All results are single runs with no standard deviations or significance tests. Many differences in Table 4.2 are small (e.g., 54.8 vs 56.1 for CIHP 10K; 60.7 vs 58.7 for CIHP 100K), and without repeated seeds or confidence intervals it is impossible to tell whether oversampling or point-count changes have any real effect.
  5. [Sections 3.2 and 4.2] Several components that materially affect the reported numbers are unspecified. The 'handcrafted rules' for cleaning backprojected labels are never itemized; the oversampled classes and oversampling ratios are not given; and DBSCAN (eps=0.03, min_samples=100) and kNN k=40 are reported without sensitivity analysis. These omissions prevent reproduction and assessment of which pipeline components drive accuracy.
minor comments (6)
  1. [Equation (3)] The interleaving sequence is printed as x1y1z1x2y1z2...; the second coordinate should be y2, not y1.
  2. [Section 3.1] 'The person's eyes should also face the camera' is not an operational criterion; specify the world-space convention (e.g., +z toward camera) used to define canonical views.
  3. [Table 4.1] 'torse' appears to be a typo for 'torso'.
  4. [Reference [8]] The citation for M2FP is a survey (arXiv:2301.00394), but the text attributes M2FP to Liang et al.; please cite the original Mask2Former-based parsing paper.
  5. [Reference [35]] The THuman2.1 dataset is cited as 'Function4D'; please cite the correct THuman2.0/THuman2.1 dataset paper.
  6. [Throughout] There are several typos: 'cannoni-cal', 'follwed', 'Sapienvs', and 'Morevoer' in Sections 1, 3, 5, and 6 respectively.

Circularity Check

1 steps flagged · score 6.0 of 10

The central accuracy claim rests on a circular benchmark: pseudo-GT labels from M2FP/Sapiens serve as both training targets and evaluation ground truth, so reported mIoU measures agreement with the 2D parsers, not validated part accuracy.

  1. other [Section 3.2 (label generation via multi-view parsing) and Section 4.3 / Table 4.2 (evaluation against those labels)]
    "For each vi, a parsing model is used which extracts a label image ... Per-point labels are then computed using a voting procedure, using the label most voted as the final choice."

    The same pseudo-ground-truth labeling pipeline (M2FP/Sapiens backprojection, voting, DBSCAN cleanup) generates the per-vertex labels used for both training supervision and test evaluation. The paper reports mIoU in Table 4.2 computed against these pseudo-GT labels after splitting the dataset into train/val/test, but the test labels originate from the same 2D parsing models that created the training labels. Thus the reported mIoU quantifies how well the PointTransformer reproduces M2FP/Sapiens outputs from geometry alone, not how accurate the 3D segmentation is against human-verified part labels. Systematic errors in the 2D parsers on rendered scans are baked into both the supervised signal and the evaluation reference, so high mIoU can be achieved even when true part accuracy is low.

full rationale

The paper contains no problematic self-citations or imported uniqueness claims; its label-generation pipeline, downsampling strategy, and PointTransformer training are described as independent components. However, the evaluation is circular with respect to the 2D parsing models: the pseudo-GT labels generated in Section 3.2 by M2FP and Sapiens are used both as training targets and as the ground truth for the mIoU metrics in Table 4.2. A random train/test split does not break this circularity because both splits inherit the same label source and its biases. Consequently, the central accuracy claim reduces to measuring agreement with the 2D parsers' outputs, not with verifiable 3D part annotations. This is a partial circularity of the benchmark rather than a derivation that is equivalent to its inputs, so a score of 6 is appropriate.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim rests on assumptions about the transferability of 2D parsing models to rendered 3D scans, the reliability of PCA-based orientation, the locality of Morton windows, and the noise-removal behavior of DBSCAN. These are untested domain assumptions rather than derived results, and they directly affect the quality of the pseudo-GT that the evaluation uses as truth.

free parameters (6)
  • YOLOv8 alignment thresholds = nose conf <0.3, shoulder asymmetry >0.15, ear conf >0.4
    Hand-chosen criteria in Section 3.1 for deciding when orientation is corrected; no sensitivity analysis is provided.
  • DBSCAN eps and min_samples = 0.03, 100
    Section 3.2; fixed across all meshes and across the normalized coordinate space without per-mesh adaptation.
  • kNN propagation k = 40
    Section 3.2; chosen for relabeling small clusters and noise points, with no ablation.
  • Morton window size = 5000
    Section 3.3; manually chosen maximum points per window for the windowed FPS; no ablation over window sizes.
  • Part-specific oversampling ratios = not specified
    Section 4.2 states oversampling for arms, face, hair, hands, lip, teeth, tongue, but ratios are never quantified; oversampling improves 10K results and hurts 100K results in Table 4.2, suggesting regime-dependent tuning.
  • Rotation correction angles = pi/8, pi/4
    Section 3.1; fixed rotation steps in the iterative alignment procedure, chosen by hand.
assumptions (5)
  • domain assumption Rendered images of aligned THuman2.1 meshes are within the input distribution of 2D human parsing models (M2FP, Sapiens).
    Section 3.2 assumes pre-trained parsing models produce reliable labels on synthetic renderings of 3D scans; no validation against manual annotation is provided.
  • ad hoc to paper The PCA eigenvector with largest eigenvalue aligns with the body's vertical axis after mean centering.
    Section 3.1 states this assumption and acknowledges it fails for lying poses, which is why YOLOv8 correction is added; the assumption is load-bearing for the initial alignment.
  • ad hoc to paper Morton-ordered windows preserve spatial locality sufficiently for windowed FPS to approximate global FPS.
    Section 3.3 relies on elementary space-filling curve properties but provides no proof or experiment showing the downsampled set preserves semantic coverage.
  • domain assumption DBSCAN in normalized coordinate space with fixed epsilon removes noise without removing true label components.
    Section 3.2 applies fixed eps and min_samples across all meshes and label classes; no per-mesh adaptation or evaluation of denoising effects is given.
  • domain assumption Backprojection with a triangle ID buffer and depth buffering produces a unique label per vertex, and majority voting across views resolves conflicts.
    Section 3.2 describes the mechanism but acknowledges occlusions and rendering artifacts still produce noise, hence the DBSCAN refinement; the correctness of the backprojection itself is not verified against known labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Part Segmentation of Human Meshes via Multi-View Human Parsing." pith.science (2026). https://pith.science/paper/MBL4APX4

@misc{pith2026250718655,
  author       = {Pith},
  title        = {Pith review of: Part Segmentation of Human Meshes via Multi-View Human Parsing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MBL4APX4}},
  note         = {Machine review of arXiv:2507.18655}
}
read the original abstract

Recent advances in point cloud deep learning have led to models that achieve high per-part labeling accuracy on large-scale point clouds, using only the raw geometry of unordered point sets. In parallel, the field of human parsing focuses on predicting body part and clothing/accessory labels from images. This work aims to bridge these two domains by enabling per-vertex semantic segmentation of large-scale human meshes. To achieve this, a pseudo-ground truth labeling pipeline is developed for the Thuman2.1 dataset: meshes are first aligned to a canonical pose, segmented from multiple viewpoints, and the resulting point-level labels are then backprojected onto the original mesh to produce per-point pseudo ground truth annotations. Subsequently, a novel, memory-efficient sampling strategy is introduced, a windowed iterative farthest point sampling (FPS) with space-filling curve-based serialization to effectively downsample the point clouds. This is followed by a purely geometric segmentation using PointTransformer, enabling semantic parsing of human meshes without relying on texture information. Experimental results confirm the effectiveness and accuracy of the proposed approach. Project code and pre-processed data is available at https://github.com/JamesMcCullochDickens/Human3DParsing/tree/master.

Figures

Figures reproduced from arXiv: 2507.18655 by the authors.

Figure 3
Figure 3. , an overview of the proposed approach for parsing of human mesh models. Initially a mesh [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 3.2
Figure 3.2. Human orientation adjustment using iterative pose estimation. [PITH_FULL_IMAGE:figures/full_fig_p004_3_2.png] view at source ↗
Figure 3.3
Figure 3.3. From rendered views and segmentation masks [PITH_FULL_IMAGE:figures/full_fig_p006_3_3.png] view at source ↗
Figures from the paper (2 more)
Figure 3.4
Figure 3.4. Figure 3.4: Visualizing the subsampling of human point [PITH_FULL_IMAGE:figures/full_fig_p007_3_4.png]
Figure 4.1
Figure 4.1. Figure 4.1: Qualitative examples of segmentation results [PITH_FULL_IMAGE:figures/full_fig_p008_4_1.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 30 canonical work pages

  1. [1]

    SMPL: A skinned multi- person linear model,

    M. Loper, N. Mahmood, J. Romero, G. Pons- Moll, and M. J. Black, “SMPL: A skinned multi- person linear model,” ACM Transactions on Graph- ics (Proc. SIGGRAPH Asia) , vol. 34, no. 6, pp. 248:1–248:16, 2015

    work page 2015

  2. [2]

    Expressive body capture: 3D hands, face, and body from a single image,

    G. Pavlakos, V . Choutas, N. Ghorbani, T. Bolkart, A. A. A. Osman, D. Tzionas, and M. J. Black, “Expressive body capture: 3D hands, face, and body from a single image,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), 2019, pp. 10 975–10 985

    work page 2019

  3. [3]

    Deep human parsing with active template regression,

    X. Liang, S. Liu, X. Shen, J. Yang, L. Liu, J. Dong, L. Lin, and S. Yan, “Deep human parsing with active template regression,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 37, no. 12, pp. 2402–2414, 2015

    work page 2015

  4. [4]

    Parsing clothing in fashion photographs,

    K. Yamaguchi, M. H. Kiapour, L. E. Ortiz, and T. L. Berg, “Parsing clothing in fashion photographs,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) . IEEE, 2012, pp. 3570–3577

    work page 2012

  5. [5]

    Instance-level human parsing via part grouping net- work,

    K. Gong, X. Liang, D. Zhang, X. Shen, and L. Lin, “Instance-level human parsing via part grouping net- work,” in Proceedings of the European Conference on Computer Vision (ECCV), 2018, pp. 770–785

    work page 2018

  6. [6]

    Pars- ing r-cnn: Instance-level human parsing from mask r-cnn,

    S. Yang, P. Liu, Z. Wang, and C. C. Loy, “Pars- ing r-cnn: Instance-level human parsing from mask r-cnn,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 8951–8960

    work page 2019

  7. [7]

    Towards real-time and lightweight human parsing with fully convolutional networks,

    Z. Xu, J. Lu, W. Wang, X. Li, Z. Liu, C. Ding, and L. Lin, “Towards real-time and lightweight human parsing with fully convolutional networks,” in Pro- ceedings of the European Conference on Computer Vision (ECCV), 2022, pp. 546–562

    work page 2022

  8. [8]

    Deep learn- ing technique for human parsing: A survey and out- look,

    L. Yang, W. Jia, S. Li, and Q. Song, “Deep learn- ing technique for human parsing: A survey and out- look,” arXiv preprint arXiv:2301.00394, 2023

    arXiv 2023

Show all 37 references
  1. [9]

    Per- pixel classification is not all you need for semantic segmentation,

    B. Cheng, A. G. Schwing, and A. Kirillov, “Per- pixel classification is not all you need for semantic segmentation,” 2021

    2021

  2. [10]

    Masked-attention mask transformer for universal image segmentation,

    B. Cheng, I. Misra, A. G. Schwing, A. Kirillov, and R. Girdhar, “Masked-attention mask transformer for universal image segmentation,” 2021

    2021

  3. [11]

    Sapiens: Foundation for human vision models,

    R. Khirodkar, T. Bagautdinov, J. Martinez, S. Zhaoen, A. James, P. Selednik, S. Anderson, and S. Saito, “Sapiens: Foundation for human vision models,” arXiv preprint arXiv:2408.12569, 2024

    2024 arXiv

  4. [12]

    Look into person: Self-supervised structure- sensitive learning and a new benchmark for human parsing,

    K. Gong, X. Liang, D. Zhang, X. Shen, and L. Lin, “Look into person: Self-supervised structure- sensitive learning and a new benchmark for human parsing,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 932–940

    2017

  5. [13]

    Understanding humans in crowded scenes: Deep nested adversarial learning and a new bench- mark for multi-human parsing,

    J. Zhao, J. Li, Y . Cheng, L. Zhou, T. Sim, S. Yan, and J. Feng, “Understanding humans in crowded scenes: Deep nested adversarial learning and a new bench- mark for multi-human parsing,” in Proceedings of the 16th European Conference on Computer Vision (ECCV), 2018, pp. 535–551

    2018

  6. [14]

    Detect what you can: Detect- ing and representing objects using holistic models and body parts,

    L.-C. Chen, G. Papandreou, I. Kokkinos, K. Mur- phy, and A. L. Yuille, “Detect what you can: Detect- ing and representing objects using holistic models and body parts,” inProceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2014, pp. 1971–1978

    2014

  7. [15]

    Transformers in 3d point clouds: A survey,

    D. Lu, Q. Xie, M. Wei, K. Gao, L. Xu, and J. Li, “Transformers in 3d point clouds: A survey,” arXiv preprint arXiv:2205.07417 , 2022. [Online]. Available: https://arxiv.org/abs/2205.07417

    2022 arXiv

  8. [16]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 652–660

    2017

  9. [17]

    Point- net++: Deep hierarchical feature learning on point sets in a metric space,

    C. R. Qi, L. Yi, H. Su, and L. J. Guibas, “Point- net++: Deep hierarchical feature learning on point sets in a metric space,” in Advances in Neural In- formation Processing Systems (NeurIPS), 2017, pp. 5099–5108

    2017

  10. [18]

    Dynamic graph cnn for learning on point clouds,

    Y . Wang, Y . Sun, Z. Liu, S. E. Sarma, M. M. Bron- stein, and J. M. Solomon, “Dynamic graph cnn for learning on point clouds,” in Proceedings of the ACM SIGGRAPH Asia Conference on Computer Graphics and Interactive Techniques. ACM, 2019, pp. 1–12

    2019

  11. [19]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, ˚A. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems , 2017, pp. 5998– 6008. 10

    2017

  12. [20]

    Point transformer,

    H. Zhao, L. Jiang, J. Jia, P. H. Torr, and V . Koltun, “Point transformer,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021, pp. 16 259–16 268

    2021

  13. [21]

    Point transformer v2: Grouped vector attention and partition-based pooling,

    Z. Wu, L. Jiang, H. Zhao, Y . Zhao, S. Lin, and J. Jia, “Point transformer v2: Grouped vector attention and partition-based pooling,” in Advances in Neural In- formation Processing Systems (NeurIPS), 2022

    2022

  14. [22]

    Point transformer v3: Building local and global repre- sentations on point clouds,

    H. Zhao, L. Jiang, Z. Wu, S. Lin, and J. Jia, “Point transformer v3: Building local and global repre- sentations on point clouds,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), 2023, pp. 16 842–16 852

    2023

  15. [23]

    3d seg- mentation of humans in point clouds with synthetic data,

    A. Takmaz, J. Schult, I. Kaftan, M. Akc ¸ay, B. Leibe, R. Sumner, F. Engelmann, and S. Tang, “3d seg- mentation of humans in point clouds with synthetic data,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2023

    2023

  16. [24]

    Scannet: Richly- annotated 3d reconstructions of indoor scenes,

    A. Dai, A. X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner, “Scannet: Richly- annotated 3d reconstructions of indoor scenes,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2017, pp. 5828–5839

    2017

  17. [25]

    Behave: Dataset and method for tracking human object interactions,

    B. L. Bhatnagar, X. Xie, I. Petrov, C. Sminchisescu, C. Theobalt, and G. Pons-Moll, “Behave: Dataset and method for tracking human object interactions,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2022, pp. 15 935–15 946

    2022

  18. [26]

    Egobody: Hu- man body shape and motion of interacting people from head-mounted devices,

    S. Zhang, Q. Ma, Y . Zhang, Z. Qian, T. Kwon, M. Pollefeys, F. Bogo, and S. Tang, “Egobody: Hu- man body shape and motion of interacting people from head-mounted devices,” in European Confer- ence on Computer Vision (ECCV). Springer, 2022, pp. 180–200

    2022

  19. [27]

    4d spatio- temporal convnets: Minkowski convolutional neural networks,

    C. B. Choy, J. Gwak, and S. Savarese, “4d spatio- temporal convnets: Minkowski convolutional neural networks,” in Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition (CVPR), 2019, pp. 3075–3084

    2019

  20. [28]

    Open-vocabulary semantic part segmentation of 3d human,

    K. Suzuki, B. Du, G. Krishnan, K. Chen, R. B. Li, and T. Nguyen, “Open-vocabulary semantic part segmentation of 3d human,” arXiv preprint arXiv:2502.19782, 2025

    2025 arXiv

  21. [29]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rol- land, L. Gustafson, T. Xiao, S. Whitehead, A. Clegg, J. Burton, W. Boyer, M. Belko, A. Hansel, E. Tovey, D. Perelshtein, A. Makarov, Z. Zheng, O. Groth, P. Doll ´ar, and R. Girshick, “Segment anything,” arXiv preprint arXiv:2304...

    2023 arXiv

  22. [30]

    Learning transferable visual models from natu- ral language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natu- ral language supervision,” in Proceedings of the 38th International Conference on Machine...

    2021

  23. [31]

    A purely algebraic justification of the kabsch-umeyama algorithm,

    J. Lawrence, J. Bernal, and C. Witzgall, “A purely algebraic justification of the kabsch-umeyama algorithm,” Journal of Research of the National Institute of Standards and Technology , vol. 124, p. 124028, 2019. [Online]. Available: https: //doi.org/10.6028/jres.124.028

    2019 doi

  24. [32]

    Ultralytics yolov8,

    G. Jocher, A. Chaurasia, Laughing-Q, J. Fang, AlexWang1900, Chuanqi305, and other contrib- utors, “Ultralytics yolov8,” https://github.com/ ultralytics/ultralytics, 2023, accessed: 2025-06-23

    2023

  25. [33]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, and P. Doll ´ar, “Microsoft coco: Common objects in context,” in European Conference on Computer Vision (ECCV) . Springer, 2014, pp. 740–755. [Online]. Available: https:/...

    2014 arXiv

  26. [34]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Ester, H.-P. Kriegel, J. Sander, and X. Xu, “A density-based algorithm for discovering clusters in large spatial databases with noise,” inProceedings of the Second International Conference on Knowledge Discovery and Data Mining (KDD). AAAI Press, 1996, pp. 226–231

    1996

  27. [35]

    Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors,

    T. Yu, Z. Zheng, K. Guo, P. Liu, Q. Dai, and Y . Liu, “Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors,” in IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR2021), June 2021

    2021

  28. [36]

    Fully con- volutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully con- volutional networks for semantic segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) , 2015, pp. 3431–3440

    2015

  29. [37]

    Shapenet: An information- rich 3d model repository,

    A. X. Chang, T. Funkhouser, L. Guibas, P. Han- rahan, Q. Huang, Z. Li, S. Savarese, M. Savva, S. Song, H. Su et al., “Shapenet: An information- rich 3d model repository,” arXiv preprint arXiv:1512.03012, 2015. 11

    2015 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.