Pith. sign in

REVIEW 3 major objections 5 minor 67 references

MetricHMSR:Metric Human Mesh and Scene Recovery from Monocular Images

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Single image recovers metric human and scene.

desk verdict Solid metric human mesh recovery with a clean ray-map idea and strong pose numbers, but the scene-depth SOTA claim rests on a PROX train/test mismatch that needs a same-training baseline. read the letter →

arxiv 2506.09919 v4 pith:JGJWKYR6 submitted 2025-06-11 cs.CV

Chentao Song , He Zhang , Haolei Yuan , Haozhe Lin , Jianhua Tao , Hongwen Zhang , Tao Yu This is my paper
classification cs.CV
keywords metrichumanmeshrecoveryboundingraymapmixture-of-expertsglobal3Dpositionhuman-guideddepthrefinementmonocularsingle-imagereconstructionSMPL
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

MetricHMSR tries to establish that a single monocular image is enough to recover a human body in metric scale — accurate pose, shape, and global 3D position — along with a metrically consistent 3D scene, without needing video or multiple frames. The key move is to feed the network a bounding ray map, a pixel-aligned camera-ray representation that encodes the camera intrinsics plus the cropping bounding box, converting monocular scale ambiguity into explicit metric cues. A Human Mixture-of-Experts routes image features to specialized experts so that local pose and global translation are learned in a disentangled way rather than through a coupled feature vector. The recovered metric human mesh is used as a geometric anchor to refine monocular depth, putting human and scene in the same metric coordinate frame. The paper reports state-of-the-art results on human mesh recovery and metric human–scene reconstruction, with the single-frame model matching video-based offline methods on global trajectory benchmarks.

What carries the argument

The bounding ray map: for every pixel $(u,v)$, the camera ray $d = K^{-1}[u,v,1]^\top$ is computed from the intrinsic matrix $K$ of the resized crop, where $K$ already encodes the crop offset and scale; feeding this map alongside the image gives the network explicit per-pixel metric cues about the camera and the subject's location within it. This is what carries the argument from monocular ambiguity to metric position. The second mechanism is HumanMoE, a soft mixture-of-experts Transformer with a Patch MoE and a Global MoE that routes tokens to specialized experts; routing statistics show body regions and image scales split across experts, demonstrating feature-level disentanglement of pose versus position. The third is the per-pixel affine depth correction $\hat{z}(x) = s(x)\, z_{\text{in}}(x) + b(x)$, anchored to the projected metric human mesh, which aligns the scene depth to the human.

What would settle it

Take a set of real images with known ground-truth camera intrinsics that deviate from the paper's longer-side approximation (for example, wide-angle shots or images cropped so the person is off-center), run MetricHMR, and compare the predicted global root translation to the known metric ground truth; the central claim is falsified if the translation error grows meaningfully with the focal-length miscalibration, since the metric position is regressed from the ray map built on that approximation.

Watch

Extended reading notes

Core claim

The paper's central claim is that metric-scale human mesh recovery and scene reconstruction can be done from a single image by making the network explicitly aware of camera geometry. It proposes MetricHMR, which constructs a bounding ray map from the camera intrinsics and the bounding-box crop, so that the network receives a ray for every pixel encoding both the focal length and the crop location; this supplies the metric cue that lets the model regress the SMPL pose, shape, and global translation $t_{\text{global}}$ directly. The HumanMoE decoder, with a soft mixture-of-experts layer that includes a dedicated ray expert and shared experts, separates local pose from global position at the feature level. On top of that, a human-guided metric depth refinement module projects the recovered human mesh to obtain per-pixel anchor depths and learns spatially varying affine corrections to a monocular depth map, aligning the human and the scene in one metric frame. The authors claim that this unified online model achieves state-of-the-art accuracy on 3DPW, EMDB, and RICH benchmarks, and is competitive with offline video-based methods while processing frames independently.

Load-bearing premise

The load-bearing premise is that when a photo's camera parameters are unknown, the focal length can be approximated from the longer side of the image and the principal point set to the image center; if that guess is wrong for strongly cropped or unusual images, the predicted metric position and the human-anchored depth alignment break down even if the network itself is sound.

Editorial extensions

If this is right

  • Any single image with estimated camera intrinsics yields a person's metric height, body shape, and 3D position in meters, so applications like augmented-reality placement, biomechanics, and virtual try-on no longer need calibrated multi-camera capture.
  • Because the model processes frames independently, video reconstruction can be done online with constant memory, yet the reported global-trajectory accuracy on EMDB-2 and RICH is comparable to offline batch methods.
  • The human mesh acts as a scale anchor for monocular depth, so in-the-wild images can be annotated with metric pseudo-ground-truth scenes, which the paper explicitly proposes for COCO, MPII, and AIC.
  • The bounding ray map is a general way to inject camera geometry into crop-based regression, suggesting the same input representation could transfer to other single-image metric estimation tasks beyond humans.

Reading between the lines

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

  • Not stated in the paper but directly testable: if the claimed disentanglement is real, resizing or shifting the bounding box should change global-position predictions while leaving local pose outputs nearly invariant; the reported routing statistics are consistent with this but stop short of a causal test.
  • The depth-refinement module inherits the base depth estimator's domain, so the method's scene claim is strongest in indoor settings like PROX; applying the same human-anchored correction to outdoor scenes with depth discontinuities would need an evaluation the paper does not provide.
  • A stronger end-to-end variant would replace the longer-side focal-length heuristic with the learned intrinsic estimator the paper already uses when available, tying metric accuracy to the calibration model rather than to a preprocessing approximation.
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

3 major / 5 minor

Summary. The paper proposes MetricHMSR, a single-image framework that recovers a metric SMPL human mesh (pose, shape, global translation) and a metric 3D scene. The human branch, MetricHMR, encodes camera intrinsics and bounding-box information into a bounding camera ray map, processes the cropped image and ray map with a mixture-of-experts design (Patch MoE and Global MoE), and regresses pose, shape, and translation. The scene branch initializes depth with MapAnything and refines it with a per-pixel affine correction (Eq. 6) conditioned on the reconstructed metric human mesh, trained on PROX RGB-D data. Experiments report strong results on 3DPW, EMDB-1/2, and RICH for human mesh recovery and global trajectory estimation, and an improved PROX depth metric (Table 5). The paper also introduces a synthetic focal-length-varying dataset, SynFocal, and describes a pseudo-GT annotation pipeline for COCO, MPII, and AIC.

Significance. If the claims hold, MetricHMSR would be a notable step toward single-frame metric human–scene reconstruction, with practical value for embodied AI and content creation. The human-mesh and trajectory results are plausible and generally stronger than prior online methods on 3DPW, EMDB, and RICH, which is a concrete and useful contribution. The paper's use of a bounding ray map is a sensible generalization of prior bounding-box encodings, and the ablations in Table 6 support the contribution of the ray map and MoE components. However, the headline claim of state-of-the-art 'metric human–scene reconstruction' rests on the PROX depth evaluation, which is weakened by the same-dataset training/evaluation protocol; the paper also ships no code, checkpoints, or error bars, limiting reproducibility. The local-pose and trajectory results are the more solid part of the submission.

major comments (3)
  1. [§3.4, Table 5] The depth refinement module is trained on the PROX RGB-D dataset (Sec. 4.1) and then evaluated on PROX in Table 5, while the baselines (Metric3D, Unidepth, MapAnything) are used zero-shot. This confounds the contribution of the human-anchor information with supervised adaptation to the PROX domain. Please either report a same-protocol baseline in which MapAnything (or another MMDE method) is fine-tuned on the same PROX training split, or evaluate the refined depth on a held-out dataset such as RICH or a SynFocal-like scene set, so that Table 5 can support the claimed SOTA in metric scene recovery.
  2. [§3.2] The metric translation t_global is regressed from the bounding ray map built from K (Eqs. 1–2). For images with unknown camera parameters, the paper falls back to approximating the focal length as the longer image side and setting the principal point to the image center (last paragraph of Sec. 3.2). This approximation is known to be inaccurate for strongly cropped or non-standard images, and because the ray map is the network's explicit metric cue, all absolute-position results inherit this error. Please report how often the fallback was used versus AnyCalib, and add an evaluation with perturbed intrinsics or a held-out calibration-error setting to quantify the robustness of t_global.
  3. [Sec. 4.1, Limitations paragraph] The in-the-wild training data for COCO, MPII, and AIC is generated as pseudo-GT by the authors' own pipeline (metric human meshes and scenes), as stated in the Limitations paragraph. This is a legitimate data-generation strategy, but the paper does not assess the quality of these pseudo-labels or their effect on the 3DPW/EMDB results. Please add a validation of the pseudo-GT (e.g., agreement with a small manually annotated subset, or an ablation training on BEDLAM+3DPW only) to show that the gains are not an artifact of self-generated labels.
minor comments (5)
  1. [Eqs. (5) and (7)] The loss weights (lambda_J2D, lambda_J3D, etc.) are listed but their values are never provided; please give the exact values in the main text or supplementary material.
  2. [Sec. 4.2, last paragraph of MMDE subsection] The sentence 'By incorporating human body information, our method enhances performance demonstrates superior performance' is ungrammatical and should be rewritten.
  3. [Figure 2 caption] The symbol '\oplus' is used in the figure to denote concatenation but is not defined in the caption or text; please define it.
  4. [Reproducibility] No code, trained models, or evaluation scripts are provided, which limits reproducibility of the reported numbers; the authors should consider releasing them or providing detailed training configurations.
  5. [Abstract and Sec. 1] The phrase 'state-of-the-art performance on both human mesh recovery and metric human–scene reconstruction' should be qualified, since the MMDE SOTA claim depends on the same-dataset PROX evaluation discussed in Major Comment 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the metric human-mesh outputs are externally supervised, and the depth refinement is a trained model evaluated on a held-out split; remaining concerns are benchmark-protocol and self-training issues, not derivations that reduce to their own inputs.

full rationale

No circular step is present in the claimed derivation chain. MetricHMR's global translation is not defined as a function of the bounding ray map in a way that makes the output equal to the input; the ray map is a camera-geometry encoding (Eqs. 1-2), and t_global is regressed by MLP heads under losses in Eq. 5 supervised by 3DPW/EMDB/RICH ground truth. HumanMoE routing is a learned architectural mechanism; the routing statistics (Fig. 5, 11-12) are descriptive post-hoc analyses, not used to define the losses or predictions. The depth refinement (Eq. 6) is a learned affine correction trained with depth/anchor/TV/variance losses (Eq. 7) on a 9:1 split of PROX; its reported numbers on the PROX test split are genuine held-out evaluations. The comparison is not protocol-matched: 'Ours' is fine-tuned on PROX while Metric3D/Unidepth/MapAnything are used zero-shot, so the MMDE gain is a benchmark-fairness concern rather than a circular derivation; a same-protocol fine-tuned baseline and an ablation without human anchors are missing. The pseudo-GT extension for COCO/MPII/AIC (Sec. 7.3 and Limitations) is a self-referential training-signal issue: the model generates targets that may later be used as training data, and this loop is not ablated; however, the headline mesh-recovery claims are validated externally, and the pseudo-GT passage is disclosed as future/limitation rather than used as a proof step. Citations to the authors' own prior work (Function4D, PyMAF) are data/baseline references and are not load-bearing for the main derivation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

Most of the load is carried by learned components rather than fitted physical constants. The three important non-architectural commitments are the camera-intrinsic approximation, the sufficiency of per-pixel affine depth correction, and the reliability of self-generated pseudo-GT labels. No new physical entities are introduced.

free parameters (4)
  • Loss weights in Eq. 5 (lambda_J2D, lambda_J3D, lambda_V3D, lambda_theta, lambda_beta, lambda_h) = not reported
    Hand-tuned weights for the multi-task HMR loss; no values or sensitivity analysis are provided.
  • Depth refinement loss weights in Eq. 7 (lambda_d, lambda_anchor, lambda_tv, lambda_var) = not reported
    Hand-tuned weights for the refinement loss; values are not given in the main text.
  • Number of routed image experts in HumanMoE = 4
    Selected after ablation in Table 6; 2, 8, and 32 route experts also work, indicating mild sensitivity.
  • Alpha_s and alpha_b bound scales for affine depth correction = not reported
    Supplementary Sec. 6.5 defines s = 1 + alpha_s tanh(s_tilde) and b = alpha_b median(...) tanh(b_tilde), but the numeric values are not specified.
assumptions (5)
  • domain assumption The SMPL model is a valid metric body representation with sufficient fidelity for pose, shape, and translation.
    Sec. 3.1 uses SMPL parameters as the output representation; the paper does not validate SMPL against higher-fidelity body models.
  • ad hoc to paper The camera ray map built from Eqs. (1)-(2) encodes the metric cues needed for translation estimation.
    The paper motivates the design but does not prove sufficiency; the +Ray ablation in Table 6 provides empirical support.
  • domain assumption When intrinsics are unavailable, focal length can be approximated by the longer image side and principal point by the image center.
    Sec. 3.2; this is a common approximation but can fail for strongly cropped or unusual images, directly affecting metric position.
  • ad hoc to paper A per-pixel affine transform of the input depth is sufficient to obtain metric-consistent scene depth.
    Sec. 3.4 and Supp. 6.5; this modeling choice is not derived from physical constraints on scene geometry.
  • ad hoc to paper Generated pseudo-GT human meshes and scenes for COCO, MPII, and AIC are accurate enough to train on.
    The limitations section states these datasets were extended with generated pseudo-GT, but the generation pipeline is not described and its contribution is not ablated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MetricHMSR:Metric Human Mesh and Scene Recovery from Monocular Images." pith.science (2026). https://pith.science/paper/JGJWKYR6

@misc{pith2026250609919,
  author       = {Pith},
  title        = {Pith review of: MetricHMSR:Metric Human Mesh and Scene Recovery from Monocular Images},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JGJWKYR6}},
  note         = {Machine review of arXiv:2506.09919}
}
read the original abstract

We introduce MetricHMSR, a novel framework for recovering metric human meshes and 3D scenes from a single monocular image. Existing methods struggle to recover metric scale due to monocular scale ambiguity and weak-perspective camera assumptions. Moreover, their fully coupled feature representations make it difficult to disentangle local pose from global translation, often requiring multi-stage pipelines that introduce accumulated errors. To address these challenges, we propose MetricHMR (Metric Human Mesh Recovery), which incorporates a bounding camera ray map representation to provide explicit metric cues for human reconstruction,together with a Human Mixture-of-Experts (HumanMoE) that dynamically routes image features to specialized experts, enabling the disentangled perception of local human pose and global metric position. Leveraging the recovered metric human as a geometric anchor, we further refine monocular metric depth estimation to achieve more accurate 3D alignment between humans and scenes.Comprehensive experiments demonstrate that our method achieves state-of-the-art performance on both human mesh recovery and metric human-scene reconstruction. Project Page: https://Metaverse-AI-Lab-THU.github.io/MetricHMSR.

Figures

Figures reproduced from arXiv: 2506.09919 by the authors.

Figure 1
Figure 1. MetricHMSR (Metric Human Mesh and Scene Recovery) reconstructs human pose, metric shape, global position, as well as scene geometry from a monocular image. Left: metric-consistent human–scene reconstruction. Right: globally consistent 3D trajectories obtained by applying MetricHMSR independently to each frame. Abstract We introduce MetricHMSR, a novel framework for recov￾ering metric human meshes and 3D scenes from … view at source ↗
Figure 2
Figure 2. Overview of MetricHMSR. Given a single image, the framework jointly recovers the metric human mesh and the scene. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Bounding ray map representation. Left: bounding box [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (11 more)
Figure 5
Figure 5. Figure 5: Routing heatmap of the deepest (last) MoE layer for [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Illustration of human-guided metric depth refinement. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Distribution of normalized SMPL root joints in cam [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Comparison of the metric measurement on PROX. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Illustration of the human body’s metric distance under [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Illustration of the human body’s metric position with [PITH_FULL_IMAGE:figures/full_fig_p012_10.png]
Figure 11
Figure 11. Figure 11: Expert allocation of the global MoE across different [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]
Figure 12
Figure 12. Figure 12: Expert allocation of the global MoE on different [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: Examples of SynFocal. Each column of images was [PITH_FULL_IMAGE:figures/full_fig_p014_13.png]
Figure 15
Figure 15. Figure 15: Impact of bounding box scale variation on shape and [PITH_FULL_IMAGE:figures/full_fig_p015_15.png]
Figure 16
Figure 16. Figure 16: Qualitative comparison of global motion trajectories. [PITH_FULL_IMAGE:figures/full_fig_p015_16.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 52 canonical work pages

  1. [1]

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

    Mykhaylo Andriluka, Leonid Pishchulin, Peter Gehler, and Bernt Schiele. 2d human pose estimation: New benchmark and state of the art analysis. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3686–3693, 2014. 6

    work page 2014

  2. [2]

    Multi-hmr: Multi-person whole-body hu- man mesh recovery in a single shot

    Fabien Baradel, Matthieu Armando, Salma Galaaoui, Ro- main Br ´egier, Philippe Weinzaepfel, Gr ´egory Rogez, and Thomas Lucas. Multi-hmr: Multi-person whole-body hu- man mesh recovery in a single shot. InProceedings of the European conference on computer vision, 2024. 2, 1

    work page 2024

  3. [3]

    Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023

    Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias M ¨uller. Zoedepth: Zero-shot trans- fer by combining relative and metric depth.arXiv preprint arXiv:2302.12288, 2023. 2, 3, 1

    arXiv 2023

  4. [4]

    Bedlam: A synthetic dataset of bodies exhibit- ing detailed lifelike animated motion

    Michael J Black, Priyanka Patel, Joachim Tesch, and Jin- long Yang. Bedlam: A synthetic dataset of bodies exhibit- ing detailed lifelike animated motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8726–8737, 2023. 6

    2023

  5. [5]

    https://www.blender.org

    Blender, 2025. https://www.blender.org. 3

    work page 2025

  6. [6]

    Federica Bogo, Angjoo Kanazawa, Christoph Lassner, Peter Gehler, Javier Romero, and Michael J. Black. Keep it smpl: Automatic estimation of 3d human pose and shape from a single image. InProceedings of the European Conference on Computer Vision, pages 561–578, 2016. 1

    work page 2016

  7. [7]

    Smpler-x: Scaling up expressive human pose and shape estimation.Advances in Neural In- formation Processing Systems, 36, 2024

    Zhongang Cai, Wanqi Yin, Ailing Zeng, Chen Wei, Qing- ping Sun, Wang Yanjun, Hui En Pang, Haiyi Mei, Mingyuan Zhang, Lei Zhang, et al. Smpler-x: Scaling up expressive human pose and shape estimation.Advances in Neural In- formation Processing Systems, 36, 2024. 1, 2

    work page 2024

  8. [8]

    Human3r: Everyone every- where all at once.arXiv preprint arXiv:2510.06219, 2025

    Yue Chen, Xingyu Chen, Yuxuan Xue, Anpei Chen, Yuliang Xiu, and Pons-Moll Gerard. Human3r: Everyone every- where all at once.arXiv preprint arXiv:2510.06219, 2025. 1, 3, 6, 7

    arXiv 2025

Show all 67 references
  1. [9]

    Accurate 3d body shape regression using metric and semantic attributes

    Vasileios Choutas, Lea M ¨uller, Chun-Hao P Huang, Siyu Tang, Dimitrios Tzionas, and Michael J Black. Accurate 3d body shape regression using metric and semantic attributes. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2718–2728, 2022. 6

    2022

  2. [10]

    Deepseekmoe: Towards ultimate ex- pert specialization in mixture-of-experts language models

    Damai Dai, Chengqi Deng, Chenggang Zhao, RX Xu, Huazuo Gao, Deli Chen, Jiashi Li, Wangding Zeng, Xingkai Yu, Yu Wu, et al. Deepseekmoe: Towards ultimate ex- pert specialization in mixture-of-experts language models. InProceedings of the 62nd Annual Meeting of the Associa- tion...

    2024

  3. [11]

    An image is worth 16×16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16×16 words: Transformers for image recognition ...

    2021

  4. [12]

    Tokenhmr: Advancing human mesh recov- ery with a tokenized pose representation

    Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, Yao Feng, and Michael J Black. Tokenhmr: Advancing human mesh recov- ery with a tokenized pose representation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1323–1333, 2024. 2, 3

    2024

  5. [13]

    Switch transformers: Scaling to trillion parameter models with sim- ple and efficient sparsity.Journal of Machine Learning Re- search, 23(120):1–39, 2022

    William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with sim- ple and efficient sparsity.Journal of Machine Learning Re- search, 23(120):1–39, 2022. 5

    2022

  6. [14]

    3d-front: 3d furnished rooms with layouts and semantics

    Huan Fu, Bowen Cai, Lin Gao, Ling-Xiao Zhang, Jiaming Wang, Cao Li, Qixun Zeng, Chengyue Sun, Rongfei Jia, Bin- qiang Zhao, et al. 3d-front: 3d furnished rooms with layouts and semantics. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 10933–10942,

  7. [15]

    Humans in 4d: Re- constructing and tracking humans with transformers

    Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Re- constructing and tracking humans with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14783–14794, 2023. 1, 2, 3, 6, 7

    2023

  8. [16]

    Vimoe: An empirical study of designing vision mixture-of-experts.arXiv preprint arXiv:2410.15732, 2024

    Xumeng Han, Longhui Wei, Zhiyang Dou, Zipeng Wang, Chenhui Qiang, Xin He, Yingfei Sun, Zhenjun Han, and Qi Tian. Vimoe: An empirical study of designing vision mixture-of-experts.arXiv preprint arXiv:2410.15732, 2024. 5

    2024 arXiv

  9. [17]

    Perspose: 3d human pose estima- tion with perspective encoding and perspective rotation

    Xiaoyang Hao and Han Li. Perspose: 3d human pose estima- tion with perspective encoding and perspective rotation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8110–8119, 2025. 2, 7, 1

    2025

  10. [18]

    Resolving 3d human pose ambiguities with 3d scene constraints

    Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2282– 2292, 2019. 6

    2019

  11. [19]

    Pow3r: Empowering un- constrained 3d reconstruction with camera and scene priors

    Wonbong Jang, Philippe Weinzaepfel, Vincent Leroy, Lour- des Agapito, and Jerome Revaud. Pow3r: Empowering un- constrained 3d reconstruction with camera and scene priors. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1071–1081, 2025. 3

    2025

  12. [20]

    End-to-end recovery of human shape and pose

    Angjoo Kanazawa, Michael J Black, David W Jacobs, and Ji- tendra Malik. End-to-end recovery of human shape and pose. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7122–7131, 2018. 1, 2, 4

    2018

  13. [21]

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

    Manuel Kaufmann, Jie Song, Chen Guo, Kaiyue Shen, Tian- jian Jiang, Chengcheng Tang, Juan Jos ´e Z ´arate, and Otmar Hilliges. Emdb: The electromagnetic database of global 3d human pose and shape in the wild. InProceedings of the IEEE/CVF International Conference on Computer V...

    2023

  14. [22]

    Mapanything: Universal feed-forward metric 3d re- construction.arXiv preprint arXiv:2509.13414, 2025

    Nikhil Keetha, Norman M ¨uller, Johannes Sch ¨onberger, Lorenzo Porzi, Yuchen Zhang, Tobias Fischer, Arno Knapitsch, Duncan Zauss, Ethan Weber, Nelson Antunes, et al. Mapanything: Universal feed-forward metric 3d re- construction.arXiv preprint arXiv:2509.13414, 2025. 1, 3, 6, 8

    2025 arXiv

  15. [23]

    Beyond weak perspective for monocular 3d human pose estimation

    Imry Kissos, Lior Fritz, Matan Goldman, Omer Meir, Ed- uard Oks, and Mark Kliger. Beyond weak perspective for monocular 3d human pose estimation. InProceedings of the European Conference on Computer Vision, pages 541–554. Springer, 2020. 2

    2020

  16. [24]

    Muhammed Kocabas, Nikos Athanasiou, and Michael J. Black. Vibe: Video inference for human body pose and shape estimation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 5252–5262, 2020. 2

    2020

  17. [25]

    Spec: Seeing people in the wild with an estimated camera

    Muhammed Kocabas, Chun-Hao P Huang, Joachim Tesch, Lea M ¨uller, Otmar Hilliges, and Michael J Black. Spec: Seeing people in the wild with an estimated camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11035–11045, 2021. 2

    2021

  18. [26]

    Learning to reconstruct 3d human pose and shape via model-fitting in the loop

    Nikos Kolotouros, Georgios Pavlakos, Michael J Black, and Kostas Daniilidis. Learning to reconstruct 3d human pose and shape via model-fitting in the loop. InProceedings of the IEEE International Conference on Computer Vision, pages 2252–2261, 2019. 1, 2

    2019

  19. [27]

    Coin: Control-inpainting diffusion prior for human and camera motion estimation

    Jiefeng Li, Ye Yuan, Davis Rempe, Haotian Zhang, Pavlo Molchanov, Cewu Lu, Jan Kautz, and Umar Iqbal. Coin: Control-inpainting diffusion prior for human and camera motion estimation. InProceedings of the European Con- ference on Computer Vision, pages 426–446. Springer, 2024. 7

    2024

  20. [28]

    Cliff: Carrying location information in full frames into human pose and shape estimation

    Zhihao Li, Jianzhuang Liu, Zhensong Zhang, Songcen Xu, and Youliang Yan. Cliff: Carrying location information in full frames into human pose and shape estimation. InPro- ceedings of the European Conference on Computer Vision, pages 590–606, 2022. 2, 4, 1

    2022

  21. [29]

    Microsoft coco: Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll´ar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vi- sion, pages 740–755. Springer, 2014. 6

    2014

  22. [30]

    Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model.arXiv preprint arXiv:2405.04434, 2024

    Aixin Liu, Bei Feng, Bin Wang, Bingxuan Wang, Bo Liu, Chenggang Zhao, Chengqi Dengr, Chong Ruan, Damai Dai, Daya Guo, et al. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model.arXiv preprint arXiv:2405.04434, 2024. 4

    2024 arXiv

  23. [31]

    Smpl: A skinned multi- person linear model.ACM Transactions on Graphics, 34(6): 1–16, 2015

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi- person linear model.ACM Transactions on Graphics, 34(6): 1–16, 2015. 4

    2015

  24. [32]

    Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis.Communications of the ACM, 65(1):99–106, 2021. 2, 4, 1

    2021

  25. [33]

    Lea M ¨uller, Ahmed A. A. Osman, Siyu Tang, Chun-Hao P. Huang, and Michael J. Black. On self-contact and human pose. InProceedings of the IEEE/CVF Conference on Com- puter Vision and Pattern Recognition, pages 9990–9999,

  26. [34]

    Soft merging of experts with adaptive routing.arXiv preprint arXiv:2306.03745, 2023

    Mohammed Muqeeth, Haokun Liu, and Colin Raffel. Soft merging of experts with adaptive routing.arXiv preprint arXiv:2306.03745, 2023. 2, 4

    2023 arXiv

  27. [35]

    Priyanka Patel and Michael J. Black. Camerahmr: Aligning people with perspective. In2025 International Conference on 3D Vision, 2025. 2, 7, 3

    2025

  28. [36]

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

    Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2019

  29. [37]

    Sharpdepth: Sharpening metric depth predictions using diffusion distillation.arXiv preprint arXiv:2411.18229, 2024

    Duc-Hai Pham, Tung Do, Phong Nguyen, Binh-Son Hua, Khoi Nguyen, and Rang Nguyen. Sharpdepth: Sharpening metric depth predictions using diffusion distillation.arXiv preprint arXiv:2411.18229, 2024. 3, 1

    2024 arXiv

  30. [38]

    Unidepth: Universal monocular metric depth estimation

    Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106–10116, 2024. 3, 8, 1

    2024

  31. [39]

    Neural localizer fields for continuous 3d human pose and shape estimation

    Istv ´an S ´ar´andi and Gerard Pons-Moll. Neural localizer fields for continuous 3d human pose and shape estimation. Advances in Neural Information Processing Systems, 37: 140032–140065, 2025. 2

    2025

  32. [40]

    Outra- geously large neural networks: The sparsely-gated mixture- of-experts layer.arXiv preprint arXiv:1701.06538, 2017

    Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outra- geously large neural networks: The sparsely-gated mixture- of-experts layer.arXiv preprint arXiv:1701.06538, 2017. 5

    2017 arXiv

  33. [41]

    World-grounded human motion recovery via gravity-view coordinates

    Zehong Shen, Huaijin Pi, Yan Xia, Zhi Cen, Sida Peng, Zechen Hu, Hujun Bao, Ruizhen Hu, and Xiaowei Zhou. World-grounded human motion recovery via gravity-view coordinates. InSIGGRAPH Asia 2024 Conference Papers, pages 1–11, 2024. 2, 3, 7

    2024

  34. [42]

    Wham: Reconstructing world-grounded humans with accu- rate 3d motion

    Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J Black. Wham: Reconstructing world-grounded humans with accu- rate 3d motion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2070– 2080, 2024. 2, 3, 7

    2024

  35. [43]

    Putting people in their place: Monocular regression of 3d people in depth

    Yu Sun, Wu Liu, Qian Bao, Yili Fu, Tao Mei, and Michael J Black. Putting people in their place: Monocular regression of 3d people in depth. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 13243–13252, 2022. 3

    2022

  36. [44]

    Trace: 5d temporal regression of avatars with dynamic cam- eras in 3d environments

    Yu Sun, Qian Bao, Wu Liu, Tao Mei, and Michael J Black. Trace: 5d temporal regression of avatars with dynamic cam- eras in 3d environments. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8856–8866, 2023. 2, 3, 7

    2023

  37. [45]

    Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras.Advances in neu- ral information processing systems, 34:16558–16569, 2021

    Zachary Teed and Jia Deng. Droid-slam: Deep visual slam for monocular, stereo, and rgb-d cameras.Advances in neu- ral information processing systems, 34:16558–16569, 2021. 3

    2021

  38. [46]

    Joachim Tesch, Giorgio Becherini, Prerana Achar, Anasta- sios Yiannakidis, Muhammed Kocabas, Priyanka Patel, and Michael J. Black. BEDLAM2.0: Synthetic humans and cam- eras in motion. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems Datasets and Be...

    2025

  39. [47]

    Anycalib: On- manifold learning for model-agnostic single-view camera calibration

    Javier Tirado-Gar ´ın and Javier Civera. Anycalib: On- manifold learning for model-agnostic single-view camera calibration. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8044–8055, 2025. 4

    2025

  40. [48]

    Recovering ac- curate 3d human pose in the wild using imus and a moving camera

    Timo V on Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering ac- curate 3d human pose in the wild using imus and a moving camera. InProceedings of the European conference on com- puter vision, pages 601–617, 2018. 6

    2018

  41. [49]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5294–5306, 2025. 1, 2, 3, 5

    2025

  42. [50]

    Continuous 3d perception model with persistent state

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10510–10522, 2025. 3

    2025

  43. [51]

    Blade: Single-view body mesh learning through accurate depth es- timation.arXiv preprint arXiv:2412.08640, 2024

    Shengze Wang, Jiefeng Li, Tianye Li, Ye Yuan, Henry Fuchs, Koki Nagano, Shalini De Mello, and Michael Stengel. Blade: Single-view body mesh learning through accurate depth es- timation.arXiv preprint arXiv:2412.08640, 2024. 2, 3

    2024 arXiv

  44. [52]

    Zolly: Zoom focal length correctly for perspective- distorted human mesh reconstruction

    Wenjia Wang, Yongtao Ge, Haiyi Mei, Zhongang Cai, Qing- ping Sun, Yanjun Wang, Chunhua Shen, Lei Yang, and Taku Komura. Zolly: Zoom focal length correctly for perspective- distorted human mesh reconstruction. InProceedings of the IEEE/CVF International Conference on Computer V...

    2023

  45. [53]

    Tram: Global trajectory and motion of 3d humans from in- the-wild videos

    Yufu Wang, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Tram: Global trajectory and motion of 3d humans from in- the-wild videos. InProceedings of the European Conference on Computer Vision, pages 467–487. Springer, 2024. 2, 3, 4, 7, 1

    2024

  46. [54]

    Prompthmr: Promptable human mesh recovery

    Yufu Wang, Yu Sun, Priyanka Patel, Kostas Daniilidis, Michael J Black, and Muhammed Kocabas. Prompthmr: Promptable human mesh recovery. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1148–1159, 2025. 2, 7

    2025

  47. [55]

    Ai challenger: A large-scale dataset for going deeper in image understanding.arXiv preprint arXiv:1711.06475, 2017

    Jiahong Wu, He Zheng, Bo Zhao, Yixin Li, Baoming Yan, Rui Liang, Wenjia Wang, Shipei Zhou, Guosen Lin, Yanwei Fu, et al. Ai challenger: A large-scale dataset for going deeper in image understanding.arXiv preprint arXiv:1711.06475, 2017. 6

    2017 arXiv

  48. [56]

    Vit- pose: Simple vision transformer baselines for human pose estimation.Advances in neural information processing sys- tems, 35:38571–38584, 2022

    Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. Vit- pose: Simple vision transformer baselines for human pose estimation.Advances in neural information processing sys- tems, 35:38571–38584, 2022. 4

    2022

  49. [57]

    Depth any- thing v2.Advances in Neural Information Processing Sys- tems, 37:21875–21911, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2.Advances in Neural Information Processing Sys- tems, 37:21875–21911, 2024. 2, 3, 1

    2024

  50. [58]

    Decoupling human and camera motion from videos in the wild

    Vickie Ye, Georgios Pavlakos, Jitendra Malik, and Angjoo Kanazawa. Decoupling human and camera motion from videos in the wild. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 21222–21232, 2023. 2, 3, 7

    2023

  51. [59]

    Metric3d: Towards zero-shot metric 3d prediction from a single image

    Wei Yin, Chi Zhang, Hao Chen, Zhipeng Cai, Gang Yu, Kaixuan Wang, Xiaozhi Chen, and Chunhua Shen. Metric3d: Towards zero-shot metric 3d prediction from a single image. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 9043–9053, 2023. 3, 8, 1

    2023

  52. [60]

    Function4d: Real-time human vol- umetric capture from very sparse consumer rgbd sensors

    Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qiong- hai Dai, and Yebin Liu. Function4d: Real-time human vol- umetric capture from very sparse consumer rgbd sensors. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 5746–5756, 2021. 3

    2021

  53. [61]

    Glamr: Global occlusion-aware human mesh recov- ery with dynamic cameras

    Ye Yuan, Umar Iqbal, Pavlo Molchanov, Kris Kitani, and Jan Kautz. Glamr: Global occlusion-aware human mesh recov- ery with dynamic cameras. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11038–11049, 2022. 7

    2022

  54. [62]

    Not all tokens are equal: Human-centric visual analysis via token clustering transformer

    Wang Zeng, Sheng Jin, Wentao Liu, Chen Qian, Ping Luo, Wanli Ouyang, and Xiaogang Wang. Not all tokens are equal: Human-centric visual analysis via token clustering transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11101– 1...

    2022

  55. [63]

    Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop

    Hongwen Zhang, Yating Tian, Xinchi Zhou, Wanli Ouyang, Yebin Liu, Limin Wang, and Zhenan Sun. Pymaf: 3d human pose and shape regression with pyramidal mesh alignment feedback loop. InProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, pages 11446–11456,

  56. [64]

    Pymaf-x: To- wards well-aligned full-body model regression from monoc- ular images.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 45(10):12287–12303, 2023

    Hongwen Zhang, Yating Tian, Yuxiang Zhang, Mengcheng Li, Liang An, Zhenan Sun, and Yebin Liu. Pymaf-x: To- wards well-aligned full-body model regression from monoc- ular images.IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, 45(10):12287–12303, 2023. 1, 2

    2023

  57. [65]

    Metric from human: Zero-shot monoc- ular metric depth estimation via test-time adaptation

    Yizhou Zhao, Hengwei Bian, Kaihua Chen, Pengliang Ji, Liao Qu, Shao-yu Lin, Weichen Yu, Haoran Li, Hao Chen, Jun Shen, et al. Metric from human: Zero-shot monoc- ular metric depth estimation via test-time adaptation. In The Thirty-eighth Annual Conference on Neural Information...

    2024

  58. [66]

    Camera, Image and Metrics In human mesh recovery, full perspective projection [2, 17, 28, 53] has gained increasing attention

    More Implementation Details 6.1. Camera, Image and Metrics In human mesh recovery, full perspective projection [2, 17, 28, 53] has gained increasing attention. The development of monocular metric depth estimation [3, 15, 37, 38, 57, 59, 65] and single-image intrinsic parameter...

  59. [67]

    SynFocal dataset

    Additional Results 7.1. SynFocal dataset. Across images with varying camera parameters, Met- ricHMSR consistently achieves accurate metric distance perception. Recognizing that existing datasets predomi- nantly feature fixed focal lengths, we construct SynFocal, a synthetic da...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.