Pith. sign in

REVIEW 4 major objections 6 minor 50 references

A new 6-million-frame 4K multi-view mocap dataset, Human4K, with Vicon-accurate SMPL-X labels, consistently improves whole-body 3D reconstruction on standard benchmarks, with the largest gains for hands, feet, and depth-ambiguous poses.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A new dataset of six million 4K multi-view frames with Vicon-mocap-derived SMPL-X annotations improves whole-body 3D human reconstruction when added to public training data.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Credible dataset contribution that fills a real gap, but the hand-label accuracy rests on a temporal sync assumption and a fitting residual that the paper never validates on held-out data. the 4 major comments →

arxiv 2607.13646 v1 pith:PJU6PKOA submitted 2026-07-15 cs.CV cs.AI

Human4K: A Large-Scale 4K Multi-View Mocap Dataset for Whole-Body 3D Human Reconstruction

Tianshun Han , Ziyu Shi , Lijian Liu , Ajian Liu , Benjia Zhou , Hugo Jair Escalante , Yanyan Liang , Sergio Escalera
show 2 more authors
Zhen Lei Jun Wan
This is my paper
classification cs.CV cs.AI
keywords Human4K3D human reconstructionSMPL-Xmotion capturemulti-view 4Kwhole-body pose estimationdatasethand reconstruction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Human4K is a large-scale dataset of over six million synchronized 4K images captured by eight cameras while professional actors perform highly articulated, self-occluding motions, with ground-truth SMPL-X body, hand, and face parameters derived from a Vicon motion-capture system. The paper's central claim is that training existing SMPL-X reconstruction methods on this dataset, in addition to public datasets, consistently reduces body and hand reconstruction errors on benchmarks like EHF and 3DPW, and that the 4K resolution and mocap-grade extremity annotations are the key reasons. If true, Human4K would fill a gap in real-world training data for whole-body reconstruction, especially for the hands and feet that pseudo-3D and synthetic datasets annotate poorly.

Core claim

The paper introduces Human4K, a dataset of 6,007,290 synchronized 4K frames captured by eight cameras at up to 15 FPS, paired with 120 FPS Vicon markers. A Motion-Retargeting and Refinement Module (MRRM) converts the Vicon BVH motion into SMPL-X parameters, including a dedicated hand-fitting stage that reduces hand joint error from 25.5 mm to 12.5 mm MPJPE. In experiments, training three representative methods (Hand4Whole, OSX-b, SMPLer-X-b) on Human4K together with public datasets (P+Ours) improves results on EHF and 3DPW: for example, Hand4Whole's EHF overall MPVPE drops from 79.00 mm to 70.10 mm and OSX-b's from 81.29 mm to 72.31 mm, with hand errors falling by roughly 4–12 mm. The paper

What carries the argument

The Motion-Retargeting and Refinement Module (MRRM) is the central mechanism: it retargets Vicon BVH skeletons to the SMPL-X kinematic tree through skeleton mapping, bone proportion scaling, rotation transfer, and canonical-space extraction, followed by an Adam-optimized hand-fitting stage that treats hand pose parameters as learnable and aligns SMPL-X hand joints to Vicon markers. The timestamp alignment between 15 FPS cameras and 120 FPS Vicon frames is the load-bearing synchronization step, and camera extrinsics are recovered via 2D-3D joint alignment with solvePnP.

Load-bearing premise

The temporal pairing of each 15 FPS camera frame with the nearest 120 FPS Vicon frame, without interpolation, is assumed accurate enough to serve as ground truth; during fast hand and foot motions, the up to ~33 ms offset can displace joints by several centimeters, and this assumption is especially critical because the headline improvements are in hands and feet.

What would settle it

Take a fast dance or sporting sequence from Human4K, re-synchronize the Vicon stream with the camera frames using higher-order interpolation or a hardware sync pulse, and re-run one of the benchmark training experiments; if the hand or foot MPVPE does not improve more than the reported training variance, then the nearest-neighbor alignment is not a limiting factor. Alternatively, compare Human4K's fitted SMPL-X hand joints against an independent high-frame-rate hand tracker on a subset: if discrepancies approach the reported 12.5 mm fitting error, the labels are too noisy to support the claime

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the reported gains hold, Human4K becomes a high-value auxiliary training set for SMPL-X whole-body reconstruction, reducing errors especially for hands and depth-ambiguous limbs across both multi-stage and one-stage architectures.
  • The dataset can serve as an in-domain benchmark that is more challenging than existing real-world sets, exposing failure modes in current models that standard training data does not cover.
  • The 4K resolution finding implies that pushing input resolution to 4K, not just adding more data, is a practical path to improving fine-grained hand and face reconstruction.
  • The MRRM pipeline may be reusable for converting other Vicon or even in-the-wild motion-capture data into SMPL-X annotations, broadening the supply of high-quality supervision beyond Human4K.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial extension: The gain pattern—large for hands and feet, smaller for body—suggests that Human4K's value lies in the quality of extremity annotations rather than in global pose diversity; a dataset with equally accurate hands but lower resolution might achieve a substantial share of the benefit.
  • Editorial extension: The resolution ablation (1K vs 2K vs 4K) is confounded by the annotation process; higher resolution may improve reconstruction primarily because the camera calibration and hand-fitting steps become more precise, not because the pixels themselves carry more information—a reader should not assume the two effects are separable.
  • Editorial extension: A testable extension would be to train on Human4K with synthetic or pseudo-3D data mixed in, to see whether the mocap-fidelity signal remains beneficial when the distribution is heavily tilted toward noisy labels.
  • Editorial extension: The documented hand MPJPE of 12.5 mm after fitting is an upper bound on ground-truth accuracy; benchmarks that report sub-10 mm hand errors after training on Human4K may be chasing annotation noise, so gains should be interpreted relative to this fitting residual.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper introduces Human4K, a large-scale multi-view 4K motion-capture dataset for whole-body 3D human reconstruction. Data are captured with eight synchronized 4K cameras and a 120 FPS Vicon system; raw mocap is retargeted to SMPL-X via a Motion-Retargeting and Refinement Module (MRRM), with a hand-fitting stage and DECA-based facial parameters. The manuscript also describes a virtual-clothing augmentation applied to half of the data. Experiments train three SMPL-X-based methods (Hand4Whole, OSX-b, SMPLer-X-b) on public data, on Human4K alone, and on the combined set, and evaluate on EHF and 3DPW, plus an in-domain Human4K test split. The central claim is that adding Human4K consistently improves whole-body reconstruction, especially for hands, feet, and depth-ambiguous poses.

Significance. If the dataset and annotation quality hold up, Human4K would be a valuable real-world complement to existing SMPL-X training corpora. Its strengths include: evaluation on two external benchmarks (EHF, 3DPW), which avoids circularity; consistent improvements across three architecturally different methods; a detailed, reproducible pipeline description; and a focus on high-resolution imagery and extremity supervision, which are known weaknesses of existing datasets. The manuscript also provides a quantitative resolution study (Table VI) and pose-distribution analysis (Fig. 8). However, the reliability of the annotation ground truth is not yet demonstrated to the level required by the paper's 'mocap-accurate' claim; two load-bearing assumptions — temporal synchronization and hand retargeting fidelity — lack validation. The dataset and code are not currently available, so the technical claims can only be judged from the manuscript.

major comments (4)
  1. [§III.A.3 (Timestamp Alignment)] The synchronization procedure pairs each 15 FPS RGB frame with the temporally closest 120 FPS Vicon frame, with no interpolation. This yields a worst-case temporal offset of ~33 ms. For fast hand/dance motions, which the dataset explicitly targets, a fingertip can move several centimeters in this window. Since the headline gains in Table IV are concentrated in hand and foot metrics, this pairing is load-bearing for the claim of 'mocap-accurate' annotations. The paper provides no quantitative validation of the alignment error, e.g., reprojection error on a held-out set, comparison against interpolated poses, or a histogram of actual offsets. Please add such an analysis, and if the offset is nontrivial, discuss its effect on annotation error.
  2. [§III.B.3 (Hand Pose Fitting; Table II)] Table II reports that after hand-fitting, the hand MPJPE is 12.5 mm (PA-MPJPE 8.4 mm). However, this is a fit residual evaluated on the same mocap data used to optimize the hand parameters. It measures how well the SMPL-X hand model can be made to match the Vicon hand joints, not the accuracy of the fitted hand poses as ground truth for an independent reconstructor. The paper itself refers to the final parameters as 'pseudo-ground-truth' in this section. Because the hand labels are central to the claimed improvements, please provide a held-out validation: e.g., compare against manually annotated 2D hand keypoints in a subset of views, evaluate a strong hand pose estimator on held-out frames, or report per-joint errors (fingertips vs. proximal joints). Without such evidence, the hand-annotation accuracy remains unverified.
  3. [Table III / Abstract] The scale of the dataset is stated ambiguously. The abstract says 'over six million 4K images'; §III.A says '6,007,290 synchronized 4K frames'; Table I lists '#Frames' as 6M; and Table III sums to 6,007,290 across scenarios. If these are per-view images, then the number of synchronized multi-view time instants is about 751K (6,007,290 / 8). If these are time instants, the total number of images is about 48M. The distinction is essential for comparing with prior datasets in Table I and for assessing the claimed scale. Please clarify the unit used in Table III and define whether a 'frame' is one camera view or one synchronized multi-view set.
  4. [§IV.B (Table IV)] The central empirical claim — 'training with Human4K consistently improves whole-body reconstruction' — rests on Table IV, but no error bars or repeated runs are reported. Some improvements are relatively modest (e.g., SMPLer-X-b EHF hand MPVPE 79.98→71.24 mm; EHF face PA-MPVPE 15.95→12.11 mm). Because training stochasticity can easily produce differences of this magnitude, the claim of consistency would be strengthened by reporting mean±std over at least two independent training runs, or at least for the main P+Ours comparison. This is not meant to detract from the consistent qualitative trend, but to establish that the reported improvements are not within noise.
minor comments (6)
  1. [Table I] The '#Frames' column is ambiguous for multi-view datasets. For Human4K, clarify whether the 6M number refers to images or synchronized multi-view time instants, and do the same for other entries (e.g., MVHumanNet, DNA-Rendering) to enable fair comparison.
  2. [§III.C] The virtual clothing augmentation is described as using 'Google's Nano model' with a generic link. Please provide the exact model version and a citation; also report a quantitative ablation showing the effect of the clothing augmentation on downstream performance, since the paper claims it 'improves robustness.'
  3. [Fig. 2] Fig. 2 shows projected Vicon joints and states they 'align well' with image observations. Please provide quantitative calibration errors (e.g., mean reprojection error in pixels) for the eight cameras, since the entire annotation pipeline depends on camera extrinsics.
  4. [§IV.E] The resolution study (Table VI) is performed only for OSX-b. Including a second method would strengthen the conclusion that 4K input benefits fine-grained reconstruction across architectures.
  5. [References] Minor typographical issues: reference [35] is inconsistently formatted ('Humaneva'), and reference [39] appears to be mislabeled ('Easy Rocap' vs. 'ZJU-Mocap' in the text). Please correct.
  6. [Dataset availability and ethics] The footnote says Human4K 'will be released when this paper is accepted.' For a dataset paper, please state the planned license and access conditions. Also mention whether informed consent and privacy/ethics approval were obtained for the 11 actors, as is standard for human-capture datasets.

Circularity Check

1 steps flagged

Central benchmark improvements are externally validated; only the hand-fit residual is presented as an accuracy measure, a minor self-referential step.

specific steps
  1. fitted input called prediction [Section III.B.3 (Hand Pose Fitting), Table II]
    "During optimization, the SMPL-X model generates 3D hand joint positions from the current pose estimates, which are then supervised using the corresponding Vicon joint locations. We employ the Mean Per Joint Position Error (MPJPE) as the objective function, minimizing the Euclidean distance between predicted and ground-truth joints ... After optimization, the hand MPJPE is reduced to 12.5 mm and the PA-MPJPE to 8.4 mm, indicating a much closer correspondence between the reconstructed and captured hand motions."

    The 12.5 mm hand MPJPE is the optimized value of the same MPJPE objective used to fit the hand pose parameters. Reporting this residual as evidence of 'mocap-accurate' hand annotation is not an independent validation: the metric is minimized on the same data, so the number is forced by construction. The paper explicitly calls the final SMPL-X parameters 'pseudo-ground-truth' in the same section, undercutting the abstract's 'mocap-accurate' claim for hands. This is a fit residual, not a held-out accuracy measure.

full rationale

Human4K's headline claim—that training on Human4K improves whole-body reconstruction on standard benchmarks—is tested on external EHF and 3DPW data. Those improvements are not forced by the annotation pipeline and therefore are not circular. No load-bearing self-citation, uniqueness theorem, or ansatz-smuggled-by-citation appears in the paper. The main circularity-adjacent issue is Table II: the hand MPJPE after MRRM hand-fitting is the optimization objective itself, so citing it as evidence of hand-label accuracy is self-referential. However, the paper is transparent that this evaluation is on the Human4K mocap data and does not claim it is a learned-model prediction. The 15 FPS/120 FPS nearest-neighbor synchronization and unvalidated hand retargeting are genuine correctness risks, but they are external accuracy concerns, not circularity in the derivation chain. The 'pseudo-ground-truth' terminology likewise flags a limitation but does not make the cross-dataset experiments circular. Overall, the central contribution is empirically benchmarked and only a minor fit-residual metric is used as an accuracy statement, so the circularity score is 2.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 0 invented entities

The central claims rest on treating Vicon marker trajectories as ground truth, assuming 15 FPS/120 FPS nearest-neighbor pairing is tight enough for hand/finger annotations, trusting DECA for face pseudo-GT, and assuming virtual clothing augmentation preserves geometry. None of these are proven in the paper; they are domain assumptions. No new physical or virtual entities are invented.

free parameters (4)
  • SMPL-X body shape β_b (10D) = Initialized to average template; minor adjustments allowed during hand-fitting
    Body shape is not measured per subject; annotations inherit the template default, which may bias body-shape supervision.
  • Hand pose θ_lh, θ_rh (15×3 each per hand) = Adam lr=1e-2, 3000 iterations, MPJPE objective
    Hand labels are the result of an optimization fit to Vicon joints; Table II reports the final fit error on the fitting data, not independent accuracy.
  • Retarget alignment scale s and bone-direction matrix M = Estimated per skeleton from T-pose
    Used to map Vicon skeleton lengths/directions to SMPL-X; errors propagate into all body pose annotations.
  • Facial parameters from DECA = DECA predictions (ψ∈R10, θf∈R3)
    Facial expression pseudo-GT is predicted by an off-the-shelf model, not mocap-measured.
axioms (6)
  • domain assumption Vicon marker motion is accurate ground truth for 3D joints, including hands
    §III.A/§III.B: The entire annotation pipeline and claimed annotation accuracy rest on trusting Vicon joint trajectories as ground truth; no independent accuracy validation beyond fit residuals.
  • domain assumption Nearest-neighbor temporal pairing between 15 FPS camera frames and 120 FPS mocap frames provides frame-level synchronization
    §III.A.3: worst-case offset ≈33 ms; for fast hand/finger motion this can be significant; no interpolation or synchronization error analysis.
  • domain assumption Vicon skeleton can be retargeted to SMPL-X through T-pose correspondence and bone alignment
    §III.B.2: if correspondences are wrong (especially 21 body joints and 15 hand joints), annotations are biased.
  • domain assumption SMPL-X body model is a sufficient anatomical model to represent captured motion
    §III.B: The paper adopts SMPL-X as standard; residual mismatch (e.g., soft tissue, hand shape) is absorbed by optimization.
  • domain assumption DECA FLAME predictions are adequate pseudo-ground-truth for face expression and jaw
    §III.B.4: facial annotation is not mocap-measured; relies on DECA.
  • domain assumption Virtual clothing augmentation preserves identity, pose, and body shape
    §III.C: augmentation uses Nano model and 5 categories; no quantitative check that geometry/pose is unchanged.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Human4K: A Large-Scale 4K Multi-View Mocap Dataset for Whole-Body 3D Human Reconstruction." pith.science (2026). https://pith.science/paper/PJU6PKOA

@misc{pith2026260713646,
  author       = {Pith},
  title        = {Pith review of: Human4K: A Large-Scale 4K Multi-View Mocap Dataset for Whole-Body 3D Human Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PJU6PKOA}},
  note         = {Machine review of arXiv:2607.13646}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent advances in 3D human reconstruction have improved overall performance, yet current models still fail in the most challenging real-world scenarios. They often produce unstable geometry, inaccurate limb articulation and unreliable predictions under depth ambiguity or self-occlusion. A key reason is that existing datasets still lack the combination of high-resolution images, high-precision annotations and diverse whole-body motions required to support robust reconstruction. To address this gap, we present Human4K, a large-scale 4K multi-view whole-body human reconstruction dataset with mocap-accurate SMPL-X annotations. Human4K contains over six million 4K images captured by an eight-view high-resolution camera system synchronized with a professional Vicon motion capture setup, covering 11 subjects performing complex, highly articulated and strongly self-occluded full-body motions. All sequences are processed by a Motion-Retargeting and Refinement Module (MRRM) to ensure precise alignment for the full body and extremities. Experimental results show that training with Human4K consistently improves whole-body reconstruction on standard benchmarks, with particularly large gains for hands, feet and depth-ambiguous limb configurations.

Figures

Figures reproduced from arXiv: 2607.13646 by Ajian Liu, Benjia Zhou, Hugo Jair Escalante, Jun Wan, Lijian Liu, Sergio Escalera, Tianshun Han, Yanyan Liang, Zhen Lei, Ziyu Shi.

Figure 1
Figure 1. Figure 1: Overview of the Human4K capture setup. (a) Eight 4K cameras [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Projected 3D Vicon joints overlaid on 2D image observations, [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pipeline for motion data acquisition and SMPL-X pose retargeting. Motion data can be obtained from a Vicon mocap system (used in Human4K) or from in-the-wild video via 3D pose estimation, converted to BVH format, retargeted to the SMPL-X skeleton, and further refined for accurate full-body and hand annotations. (a) (b) (c) (d) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of hand pose quality before and after fitting. (a) [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Examples of virtual clothing augmentation applied to Human4K. Each [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sample frames from the Human4K dataset, illustrating diverse motions [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Effect of Human4K on challenging cases. For each sample, we show the RGB input, the reconstruction of models trained only on public datasets, and the reconstruction after adding Human4K. Human4K notably improves performance under depth ambiguity and strong limb occlusion. (P) perform poorly on Human4K, with large errors across both MPVPE and MPJPE metrics. This performance drop highlights that Human4K is s… view at source ↗
Figure 8
Figure 8. Figure 8: The distribution of human body pose data from different datasets after [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: A sequence of human body poses captured from different angles (front, [PITH_FULL_IMAGE:figures/full_fig_p009_9.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 2 linked inside Pith

  1. [1]

    An integrated platform for live 3D human reconstruction and motion capturing

    Dimitrios S Alexiadis et al. “An integrated platform for live 3D human reconstruction and motion capturing”. In:IEEE Transactions on Circuits and Systems for Video Technology27.4 (2016), pp. 798–813

    2016

  2. [2]

    2d human pose estimation: New benchmark and state of the art analysis

    Mykhaylo Andriluka et al. “2d human pose estimation: New benchmark and state of the art analysis”. In:Pro- ceedings of the IEEE Conference on computer Vision and Pattern Recognition. 2014, pp. 3686–3693

    2014

  3. [3]

    Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion

    Michael J Black et al. “Bedlam: A synthetic dataset of bodies exhibiting detailed lifelike animated motion”. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 8726–8737

    2023

  4. [4]

    Humman: Multi-modal 4d human dataset for versatile sensing and modeling

    Zhongang Cai et al. “Humman: Multi-modal 4d human dataset for versatile sensing and modeling”. In:Euro- pean Conference on Computer Vision. Springer. 2022, pp. 557–577

    2022

  5. [5]

    Playing for 3d human recovery

    Zhongang Cai et al. “Playing for 3d human recovery”. In:arXiv preprint arXiv:2110.07588(2021)

    Pith/arXiv arXiv 2021

  6. [6]

    Zhongang Cai et al.SMPLer-X: Scaling Up Expressive Human Pose and Shape Estimation. 2023. arXiv: 2309. 17448[cs.CV].URL: https : / / arxiv. org / abs / 2309 . 17448

    2023

  7. [7]

    3D human body reconstruction based on SMPL model

    Dongyue Chen et al. “3D human body reconstruction based on SMPL model”. In:The Visual Computer39.5 (2023), pp. 1893–1906

    2023

  8. [8]

    Anatomy-aware 3d human pose estimation with bone-based pose decomposition

    Tianlang Chen et al. “Anatomy-aware 3d human pose estimation with bone-based pose decomposition”. In: IEEE Transactions on Circuits and Systems for Video Technology32.1 (2021), pp. 198–209

    2021

  9. [9]

    Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering

    Wei Cheng et al. “Dna-rendering: A diverse neural actor repository for high-fidelity human-centric rendering”. In:Proceedings of the IEEE/CVF International Con- ference on Computer Vision. 2023, pp. 19982–19993

    2023

  10. [10]

    Generalizable Neural Performer: Learning Robust Radiance Fields for Human Novel View Synthesis

    Wei Cheng et al. “Generalizable Neural Performer: Learning Robust Radiance Fields for Human Novel View Synthesis”. In:arXiv preprint arXiv:2204.11798 (2022)

    Pith/arXiv arXiv 2022

  11. [11]

    Learning an animatable detailed 3D face model from in-the-wild images

    Yao Feng et al. “Learning an animatable detailed 3D face model from in-the-wild images”. In:ACM Trans- actions on Graphics (ToG)40.4 (2021), pp. 1–13

    2021

  12. [12]

    PESTalk: Speech-Driven 3D Facial Animation with Personalized Emotional Styles

    Tianshun Han et al. “PESTalk: Speech-Driven 3D Facial Animation with Personalized Emotional Styles”. In: Proceedings of the 33rd ACM International Conference on Multimedia. 2025, pp. 7893–7901

    2025

  13. [13]

    PMMTalk:Speech-Driven 3D Facial Animation from Complementary Pseudo Multi- modal Features

    Tianshun Han et al. “PMMTalk:Speech-Driven 3D Facial Animation from Complementary Pseudo Multi- modal Features”. In:IEEE Transactions on Multimedia (2024)

    2024

  14. [14]

    Personalized graph generation for monocular 3D human pose and shape estimation

    Junxing Hu et al. “Personalized graph generation for monocular 3D human pose and shape estimation”. In: IEEE Transactions on Circuits and Systems for Video Technology34.4 (2023), pp. 2399–2413

    2023

  15. [15]

    Capturing and Inferring Dense Full-Body Human-Scene Contact

    Chun-Hao P. Huang et al. “Capturing and Inferring Dense Full-Body Human-Scene Contact”. In:Proceed- ings IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR). June 2022, pp. 13274–13285

    2022

  16. [16]

    Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments

    Catalin Ionescu et al. “Human3. 6m: Large scale datasets and predictive methods for 3d human sensing in natural environments”. In:IEEE transactions on pattern analysis and machine intelligence36.7 (2013), pp. 1325–1339

    2013

  17. [17]

    Clustered pose and nonlinear appearance models for human pose esti- mation

    Sam Johnson and Mark Everingham. “Clustered pose and nonlinear appearance models for human pose esti- mation.” In:bmvc. V ol. 2. 4. Aberystwyth, UK. 2010, p. 5

    2010

  18. [18]

    Emdb: The electromagnetic database of global 3d human pose and shape in the wild

    Manuel Kaufmann et al. “Emdb: The electromagnetic database of global 3d human pose and shape in the wild”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 14632– 14643

    2023

  19. [19]

    Learning a model of facial shape and expression from 4D scans

    Tianye Li et al. “Learning a model of facial shape and expression from 4D scans.” In:ACM Trans. Graph.36.6 (2017), pp. 194–1

    2017

  20. [20]

    Exploiting temporal contexts with strided transformer for 3d human pose estimation

    Wenhao Li et al. “Exploiting temporal contexts with strided transformer for 3d human pose estimation”. In: IEEE Transactions on Multimedia25 (2022), pp. 1282– 1293

    2022

  21. [21]

    Detailed 3D human body reconstruction from multi- view images combining voxel super-resolution and learned implicit representation

    Zhongguo Li, Magnus Oskarsson, and Anders Heyden. “Detailed 3D human body reconstruction from multi- view images combining voxel super-resolution and learned implicit representation”. In:Applied Intelligence 52.6 (2022), pp. 6739–6759

    2022

  22. [22]

    One-stage 3d whole-body mesh recovery with component aware transformer

    Jing Lin et al. “One-stage 3d whole-body mesh recovery with component aware transformer”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023, pp. 21159–21168. 11

    2023

  23. [23]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin et al. “Microsoft coco: Common objects in context”. In:Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13. Springer. 2014, pp. 740–755

    2014

  24. [24]

    SeIF: semantic-constrained deep implicit function for single-image 3d head reconstruc- tion

    Leyuan Liu et al. “SeIF: semantic-constrained deep implicit function for single-image 3d head reconstruc- tion”. In:IEEE Transactions on Multimedia26 (2024), pp. 10106–10120

    2024

  25. [25]

    Human motion tracking by multiple RGBD cameras

    Zhenbao Liu et al. “Human motion tracking by multiple RGBD cameras”. In:IEEE Transactions on Circuits and Systems for Video Technology27.9 (2016), pp. 2014– 2027

    2016

  26. [26]

    SMPL: A Skinned Multi-Person Linear Model

    Matthew Loper et al. “SMPL: A Skinned Multi-Person Linear Model”. In:ACM Transactions on Graphics, (Proc. SIGGRAPH Asia)34.6 (Oct. 2015), 248:1– 248:16

    2015

  27. [27]

    AMASS: Archive of Motion Capture as Surface Shapes

    Naureen Mahmood et al. “AMASS: Archive of Motion Capture as Surface Shapes”. In:International Confer- ence on Computer Vision. Oct. 2019, pp. 5442–5451

    2019

  28. [28]

    Monocular 3d human pose estimation in the wild using improved cnn supervision

    Dushyant Mehta et al. “Monocular 3d human pose estimation in the wild using improved cnn supervision”. In:2017 international conference on 3D vision (3DV). IEEE. 2017, pp. 506–516

    2017

  29. [29]

    Accurate 3D hand pose estimation for whole-body 3D human mesh estimation

    Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. “Accurate 3D hand pose estimation for whole-body 3D human mesh estimation”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022, pp. 2308–2317

    2022

  30. [30]

    Neuralannot: Neural annotator for 3d human mesh training sets

    Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. “Neuralannot: Neural annotator for 3d human mesh training sets”. In:Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. 2022, pp. 2299–2307

    2022

  31. [31]

    Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image

    Gyeongsik Moon et al. “Interhand2. 6m: A dataset and baseline for 3d interacting hand pose estimation from a single rgb image”. In:Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23– 28, 2020, Proceedings, Part XX 16. Springer. 2020, pp. 548–564

    2020

  32. [32]

    AGORA: Avatars in geography optimized for regression analysis

    Priyanka Patel et al. “AGORA: Avatars in geography optimized for regression analysis”. In:Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021, pp. 13468–13478

    2021

  33. [33]

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

    Georgios Pavlakos et al. “Expressive body capture: 3d hands, face, and body from a single image”. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019, pp. 10975–10985

    2019

  34. [34]

    Embodied Hands: Modeling and Capturing Hands and Bodies Together

    Javier Romero, Dimitrios Tzionas, and Michael J. Black. “Embodied Hands: Modeling and Capturing Hands and Bodies Together”. In:ACM Transactions on Graphics, (Proc. SIGGRAPH Asia). 245:1–245:17 36.6 (Nov. 2017)

    2017

  35. [35]

    Humaneva: Syn- chronized video and motion capture dataset for evalua- tion of articulated human motion

    Leonid Sigal and Michael J Black. “Humaneva: Syn- chronized video and motion capture dataset for evalua- tion of articulated human motion”. In:Brown Univert- sity TR120.2 (2006)

    2006

  36. [36]

    FTCM: Frequency-temporal col- laborative module for efficient 3D human pose estima- tion in video

    Zhenhua Tang et al. “FTCM: Frequency-temporal col- laborative module for efficient 3D human pose estima- tion in video”. In:IEEE Transactions on Circuits and Systems for Video Technology34.2 (2023), pp. 911–923

    2023

  37. [37]

    Learning from synthetic humans

    Gul Varol et al. “Learning from synthetic humans”. In: Proceedings of the IEEE conference on computer vision and pattern recognition. 2017, pp. 109–117

    2017

  38. [38]

    Recovering accurate 3d human pose in the wild using imus and a moving camera

    Timo V on Marcard et al. “Recovering accurate 3d human pose in the wild using imus and a moving camera”. In:Proceedings of the European conference on computer vision (ECCV). 2018, pp. 601–617

    2018

  39. [39]

    Easy Rocap: A Low-Cost and Easy- to-Use Motion Capture System for Drones

    Haoyu Wang et al. “Easy Rocap: A Low-Cost and Easy- to-Use Motion Capture System for Drones”. In:Drones 8.4 (2024), p. 137

    2024

  40. [40]

    Freeman: Towards benchmarking 3d human pose estimation under real-world conditions

    Jiong Wang et al. “Freeman: Towards benchmarking 3d human pose estimation under real-world conditions”. In:Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. 2024, pp. 21978– 21988

    2024

  41. [41]

    Prior-guided multi-view 3d head reconstruction

    Xueying Wang et al. “Prior-guided multi-view 3d head reconstruction”. In:IEEE Transactions on Multimedia 24 (2021), pp. 4028–4040

    2021

  42. [42]

    Global and local spatio-temporal encoder for 3D human pose estimation

    Yong Wang et al. “Global and local spatio-temporal encoder for 3D human pose estimation”. In:IEEE Transactions on Multimedia26 (2023), pp. 4039–4049

    2023

  43. [43]

    Mvhumannet: A large-scale dataset of multi-view daily dressing human captures

    Zhangyang Xiong et al. “Mvhumannet: A large-scale dataset of multi-view daily dressing human captures”. In:Proceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition. 2024, pp. 19801– 19811

    2024

  44. [44]

    Synbody: Synthetic dataset with lay- ered human models for 3d human perception and mod- eling

    Zhitao Yang et al. “Synbody: Synthetic dataset with lay- ered human models for 3d human perception and mod- eling”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 20282– 20292

    2023

  45. [45]

    STAF: 3D human mesh recovery from video with spatio-temporal alignment fusion

    Wei Yao et al. “STAF: 3D human mesh recovery from video with spatio-temporal alignment fusion”. In: IEEE Transactions on Circuits and Systems for Video Technology34.11 (2024), pp. 10564–10577

    2024

  46. [46]

    Pose2seg: Detection free hu- man instance segmentation

    Song-Hai Zhang et al. “Pose2seg: Detection free hu- man instance segmentation”. In:Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019, pp. 889–898

    2019

  47. [47]

    Video-based outdoor human reconstruc- tion

    Hao Zhu et al. “Video-based outdoor human reconstruc- tion”. In:IEEE Transactions on Circuits and Systems for Video Technology27.4 (2016), pp. 760–770

    2016

  48. [48]

    H3wb: Human3. 6m 3d wholebody dataset and benchmark

    Yue Zhu, Nermin Samet, and David Picard. “H3wb: Human3. 6m 3d wholebody dataset and benchmark”. In: Proceedings of the IEEE/CVF international conference on computer vision. 2023, pp. 20166–20177

    2023

  49. [49]

    Freihand: A dataset for markerless capture of hand pose and shape from sin- gle rgb images

    Christian Zimmermann et al. “Freihand: A dataset for markerless capture of hand pose and shape from sin- gle rgb images”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision. 2019, pp. 813–822

    2019

  50. [50]

    Human pose and shape estimation from single polarization images

    Shihao Zou et al. “Human pose and shape estimation from single polarization images”. In:IEEE Transactions on Multimedia25 (2022), pp. 3560–3572

    2022

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.