Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

CameraHMR: Aligning People with Perspective

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Per-image camera prediction improves 3D human pose and shape accuracy

desk verdict Solid and honestly reported: predicted camera intrinsics plus denser pseudo ground truth deliver real HPS gains, though the iterative self-training loop is never ablated so the credit attribution stays coarse. read the letter →

arxiv 2411.08128 v1 pith:BR742OIE submitted 2024-11-12 cs.CV

Priyanka Patel , Michael J. Black This is my paper · ORCID
classification cs.CV
keywords 3DhumanposeandshapeestimationcameraintrinsicsfieldofviewpredictionpseudogroundtruthSMPLfittingperspectivemodeldensesurfacekeypointsmonocularmeshrecovery
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper's central claim is that monocular 3D human pose and shape estimation is held back by the wrong camera model: training data are fitted and methods are deployed under a weak-perspective camera with default intrinsics, so perspective distortion in real photos is encoded into the pseudo ground truth as 3D pose error. The paper proposes to estimate each image's vertical field of view with a learned regressor, HumanFoV, and to use those predicted intrinsics in a full perspective camera during SMPL fitting, alongside dense surface keypoints that constrain body shape. It reports that the resulting refit pseudo ground truth and a modified HMR2.0-style regressor, CameraHMR, achieve state-of-the-art accuracy on established benchmarks, with the largest gains on datasets containing varied cameras and off-center people. If true, the work shows that per-image camera estimation is a practical and transferable ingredient for human mesh recovery, not just a refinement for calibration specialists.

What carries the argument

The load-bearing mechanism is HumanFoV: an HRNet-based regressor that maps aspect-preserving 256x256 images of people to a vertical field of view, with the focal length recovered from the image height via $f_y = H / (2 \tan(\upsilon/2))$. The asymmetric loss on the field of view encodes the observation that underestimating the field of view hurts reconstructed 3D poses less than overestimating it. HumanFoV's output drives two things: CamSMPLify, a modified SMPLify that fits SMPL under full perspective projection using both the original 17 joints and 138 dense surface keypoints; and CameraHMR, a transformer regressor whose decoder cross-attends to image tokens plus bounding box and focal length tokens. The iteration between CameraHMR and CamSMPLify is what turns the improved fitting procedure into a better regressor.

What would settle it

Measure HumanFoV's field-of-view error on a held-out set of human images with known intrinsics from varied cameras and crops, and check whether the error grows with focal length, crop amount, or off-center framing. If large FoV errors appear precisely on images with strong perspective, or if CamSMPLify fits on those images show biased poses compared with fits made using the true intrinsics, the central claim that predicted intrinsics drive the improvement would be contradicted. A simpler observational check is to run CameraHMR with per-image predicted intrinsics versus a default focal length stratified by true FoV; the claimed advantage should appear mainly at wide FoV and vanish or invert if the predictor is biased.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a full perspective camera with predicted per-image intrinsics can be integrated end to end into human pose and shape estimation, and doing so improves both the training signal and the trained model. HumanFoV regresses the vertical field of view from a 256x256 aspect-preserving crop and is trained on roughly 500K Flickr images with EXIF-derived focal lengths, using an asymmetric loss that penalizes overestimating the field of view more than underestimating it. CamSMPLify then refits SMPL to the 4DHumans images with perspective projection, adding 138 dense surface keypoints produced by a detector trained on BEDLAM, and the CameraHMR transformer consumes the focal length as an input token. Training is iterated: the current model initializes the next round of fitting, and the improved pseudo ground truth trains the next model. The paper reports the strongest results on benchmarks with varied intrinsics, such as 45.4 PA-MPJPE and 97.0 PVE on EMDB and large margins on SPEC-SYN, while retaining near-best 2D alignment on COCO.

Load-bearing premise

The load-bearing premise is that HumanFoV, trained on Flickr images with EXIF focal lengths, predicts the field of view accurately enough for arbitrary in-the-wild images of people, including cropped, portrait, and unusual-aspect-ratio images.

Editorial extensions

If this is right

  • Cameras need not be known at test time: the same HumanFoV prediction used to build training data is used at inference, so the perspective camera model is available for any single image of a person.
  • Benchmarks with wide-angle lenses, off-center subjects, or strong foreshortening are where the predicted intrinsics matter most; the ablation on SPEC-SYN shows the default focal length fails there.
  • Dense surface keypoints in the fitting loop should make body shape predictions more diverse and realistic, not just more aligned, with SSP-3D as the evaluation target.
  • The improved pseudo ground truth is reusable: the paper releases the pGT and pipeline, so other human pose and shape models can train on perspective-corrected fits without repeating the data collection.
  • Because AVA clips with horizontally stretched aspect ratios violate the single-focal-length assumption, datasets used for this pipeline must be filtered for consistent sensor geometry.

Reading between the lines

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

  • My inference: the 'body as calibration object' idea could generalize beyond people; articulated objects with known proportions could supply field-of-view cues in settings where vanishing points are absent, such as close-up animal or robot imagery.
  • My inference: the claimed gains depend on HumanFoV generalizing to crops and aspect ratios it never saw as clean EXIF images; a natural test is to measure FoV error and downstream pose error separately on heavily cropped or non-standard-ratio photos.
  • My inference: the iterative training loop is a form of self-training, so it could silently reinforce the initial model's systematic errors; comparing against pGT produced from an independent initialization would reveal whether iteration is correcting or entrenching bias.
  • My inference: if FoV prediction is accurate but the assumption of zero radial distortion and a centered principal point fails, such as on phone wide-angle lenses, the remaining distortion will be absorbed into pose and shape errors; testing on images with known distortion coefficients would bound this effect.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This paper proposes CameraHMR, a monocular 3D human pose and shape estimation method that replaces the common weak-perspective camera assumption with a full perspective camera whose intrinsics are predicted by a learned field-of-view regressor (HumanFoV). The authors further improve the pseudo ground truth (pGT) used for training by (i) fitting SMPL with the predicted intrinsics in a modified SMPLify (CamSMPLify), (ii) adding a dense surface keypoint detector (DenseKP) trained on synthetic data, and (iii) iterating model training and fitting in a self-training loop. The final model is evaluated on 3DPW, EMDB, RICH, SPEC-SYN, COCO-val, and SSP-3D, and the paper reports state-of-the-art PA-MPJPE/PVE on several benchmarks.

Significance. If the results hold, the work is significant: it addresses a recognized limitation of weak-perspective camera models in monocular HPS and presents a practical pipeline for generating higher-quality pGT for large-scale in-the-wild training data. The paper includes a new human-centric FoV dataset and a dense keypoint detector, and reports consistent improvements over strong baselines on external benchmarks, with particularly large margins on SPEC-SYN. The availability of code and pGT would facilitate reproducibility. However, the empirical attribution of the gains to the proposed components is incomplete, as detailed in the major comments.

major comments (3)
  1. [Section 3.4 and Table 4] The iterative self-training loop is not ablated, so the contribution of the proposed pGT improvements cannot be separated from the effect of the architectural changes and the iterative refinement. In Table 4, the comparison between HMR2.0b (trained on original 4DHumans) and CameraHMR (4DH) changes at least three factors simultaneously: the pGT generation (HumanFoV intrinsics plus dense keypoints), the network architecture (bounding box and focal length tokens, perspective projection), and the iterative fitting loop. I recommend adding an ablation that trains CameraHMR on the original 4DHumans pGT with the same architecture, and an ablation comparing one pass versus two passes of CamSMPLify, so that the effect of the improved pGT is isolated.
  2. [Section 3.4 and Figure 2] The claim that CamSMPLify produces 'significantly more realistic' pGT is supported only by qualitative examples and downstream performance; there is no quantitative validation of the pGT against known 3D ground truth. A direct test would be to run CamSMPLify on a dataset with ground-truth 3D bodies, such as the 3DPW training set or a held-out BEDLAM sequence, and report the fitting error (e.g., PA-MPJPE/PVE) for the original pGT versus the improved pGT. This would substantiate the causal chain in the abstract, which is currently only indirectly assessed.
  3. [Section 3.4 (Surface Keypoints) and Eq. (4)] The dense surface keypoint detector is a central component of the proposed pGT pipeline, but its accuracy is never evaluated, and no ablation removes the dense keypoint term from CamSMPLify. I suggest reporting 2D keypoint detection error for DenseKP on a held-out synthetic or real benchmark, and/or an ablation of Eq. (4) with and without the ES2d term, to show that the dense keypoints are responsible for the improved shape estimates rather than the other changes in the fitting process.
minor comments (4)
  1. [Table 1] The entry for Perspective Fields appears as '15.38.0' due to a missing space; it should read '15.3 8.0 14.0 12.8 18.0' for the five benchmark columns.
  2. [Throughout] The text uses both '4D-Humans' and '4DHumans'; please use a consistent name throughout.
  3. [Abstract] The sentence 'Code and pGT is available' should be 'Code and pGT are available' to maintain subject-verb agreement.
  4. [Conclusion] The paper does not include a limitations discussion; a short paragraph noting the dependence on HumanFoV accuracy and the synthetic-to-real gap for DenseKP would be helpful for readers.

Circularity Check

0 steps flagged · score 0.0 of 10

No material circularity: the central accuracy claim is tested against external 3D benchmarks with ground-truth pose, shape, and camera intrinsics, and the self-training loop is a confound rather than a by-construction reduction.

full rationale

The derivation chain is not circular by construction. HumanFoV is trained on Flickr images whose vertical field of view is derived from EXIF focal length via Eq. (6); these labels are independent of the subsequent HPS pipeline. CamSMPLify uses HumanFoV intrinsics plus dense surface keypoints from DenseKP (trained on BEDLAM/AGORA synthetic ground truth) to fit SMPL to 4D-Humans images, producing new pGT. CameraHMR is then trained on BEDLAM/AGORA plus this pGT, with the focal length supplied as a token. The central results in Table 4 and Table 5 are measured on 3DPW, EMDB, RICH, SPEC-SYN, and SSP-3D, all of which provide externally supplied ground truth that is not produced by the paper's own fitting loop. No equation in the paper defines a predicted quantity in terms of the quantity it is meant to explain; Eq. (1) and Eq. (6) are just the standard focal-length/FoV conversion. The iterative refinement described in Sec. 3.4 ('Iteration') is a self-training loop in the spirit of SPIN: CameraHMR v1 initializes CamSMPLify, the resulting pGT trains v2, and v2 re-initializes another fitting pass. This could confound an attribution of the final gains to the new camera model, and the paper does not ablate the loop or compare intermediate pGT fits to known 3D ground truth; that is an experimental-design/correctness risk, not a circularity, because the evaluation itself remains external. Self-citations to BEDLAM, AGORA, SPEC, and TokenHMR are citations to datasets, benchmarks, and prior methods with independent content, not to an unverified theorem that forces the paper's choice. The paper is therefore self-contained against external benchmarks, and no load-bearing reduction of a prediction to its own input is exhibited.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the reliability of the EXIF-derived FoV labels, the transferability of synthetic-trained keypoint detectors, and a set of hand-chosen fitting hyperparameters. No new physical entities are introduced.

free parameters (3)
  • CamSMPLify loss weights (lambda_S2d, lambda_J2d, lambda_beta, lambda_int) = Not specified in the paper
    These weights control the relative importance of dense keypoints, joints, shape prior, and initial-vertex regularization in Equation (4). They are chosen by hand and not justified by a separate validation procedure.
  • Threshold tau for pGT filtering = Not specified in the paper
    Samples with fitting error above tau are discarded. The choice of tau affects the size and quality of the pGT dataset and is not analyzed.
  • Asymmetric loss factor 3 in HumanFoV = 3
    The loss in Equation (2) penalizes overestimation of FoV three times more than underestimation, based on a heuristic from prior work. The factor is set by hand.
assumptions (4)
  • domain assumption EXIF FocalLengthIn35mmFormat provides accurate vertical field of view when combined with the assumed 24mm sensor height.
    Used in Section 4.1 to compute ground truth FoV for the Flickr training set. If the EXIF data is wrong or the 35mm equivalent assumption fails for certain cameras or crops, the HumanFoV labels are biased.
  • domain assumption The human body provides sufficient visual cues for field-of-view estimation from a single image.
    This premise underlies the whole HumanFoV approach. The paper argues that bodies have regular proportions, but does not prove that this cue is identifiable or reliable across diverse images.
  • domain assumption Dense surface keypoints learned on synthetic data (BEDLAM, AGORA) transfer to real in-the-wild images.
    The DenseKP detector is trained on synthetic data and applied to 4D-Humans real images. If the synthetic-to-real gap is large, the keypoint predictions could be noisy and degrade the pGT shape.
  • standard math Standard perspective projection and SMPL model are valid representations.
    The paper assumes fx = fy, principal point at image center, no distortion, and uses the SMPL body model. These are standard assumptions in the HPS literature.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CameraHMR: Aligning People with Perspective." pith.science (2026). https://pith.science/paper/BR742OIE

@misc{pith2026241108128,
  author       = {Pith},
  title        = {Pith review of: CameraHMR: Aligning People with Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BR742OIE}},
  note         = {Machine review of arXiv:2411.08128}
}
read the original abstract

We address the challenge of accurate 3D human pose and shape estimation from monocular images. The key to accuracy and robustness lies in high-quality training data. Existing training datasets containing real images with pseudo ground truth (pGT) use SMPLify to fit SMPL to sparse 2D joint locations, assuming a simplified camera with default intrinsics. We make two contributions that improve pGT accuracy. First, to estimate camera intrinsics, we develop a field-of-view prediction model (HumanFoV) trained on a dataset of images containing people. We use the estimated intrinsics to enhance the 4D-Humans dataset by incorporating a full perspective camera model during SMPLify fitting. Second, 2D joints provide limited constraints on 3D body shape, resulting in average-looking bodies. To address this, we use the BEDLAM dataset to train a dense surface keypoint detector. We apply this detector to the 4D-Humans dataset and modify SMPLify to fit the detected keypoints, resulting in significantly more realistic body shapes. Finally, we upgrade the HMR2.0 architecture to include the estimated camera parameters. We iterate model training and SMPLify fitting initialized with the previously trained model. This leads to more accurate pGT and a new model, CameraHMR, with state-of-the-art accuracy. Code and pGT are available for research purposes.

Figures

Figures reproduced from arXiv: 2411.08128 by the authors.

Figure 1
Figure 1. Putting people in perspective. In contrast to common methods like HMR2.0, CameraHMR estimates 3D human shape and pose using a perspective camera by leveraging a learned regressor, HumanFoV, to estimate the appropriate camera intrinsics. Note how this improves the estimated pose when there is strong foreshortening. Our approach exploits new pseudo ground-truth data and a new dense surface keypoint detector that impro… view at source ↗
Figure 2
Figure 2. Pseudo-Ground-Truth (pGT) training data. Row 1: example images from the 4DHumans dataset. Rows 2 and 3: original pGT overlaid and viewed from a different perspective. Rows 4 and 5: our improved pGT using CamSMPLify. Note that our approach reduces the bias towards bent knees (columns 1, 5, 6), improves 3D pose and image alignment when there is foreshortening (Column 2, 4, 7, 9), and estimates more realistic body shap… view at source ↗
Figure 3
Figure 3. Overview of CamSMPLify: The DenseKP module pro￾cesses cropped images to produce dense surface keypoints, while the HumanFoV module uses full images to estimate camera intrin￾sics. The output from these are used by CamSMPLify to optimize the SMPL model parameters, β, θ, and the global translation t full . Our iterative training strategy starts with initial estimates, Vinit from CameraHMR, which are used to regularize… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Qualitative results of different baselines on LSP [21] and MPII [4] test images. CameraHMR achieves better 3D pose and shape reconstruction while also achieving more accurate 2D alignment compared to other SOTA methods trained on comparable datasets. 8 [PITH_FULL_IMAG…
Figure 5
Figure 5. Figure 5: Focal Length distribution of images used in training Hu [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: CameraHMR results on landscape images downloaded from Pexels [2]. [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]
Figure 7
Figure 7. Figure 7: CameraHMR results on portrait images downloaded from Pexels [2]. [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: CameraHMR achieves more accurate 3D pose estimation, shape reconstruction, and 2D alignment with the image even for [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Humans as a Calibration Pattern: Dynamic 3D Scene Reconstruction from Unsynchronized and Uncalibrated Videos

    cs.CV 2024-12 conditional novelty 7.0 of 10

    Dynamic 3D scenes can be reconstructed from unsynchronized, uncalibrated multi-view videos by first aligning estimated human motion across views and then refining the alignment during neural field training.

  2. WATCH: World-aware Allied Trajectory and pose reconstruction for Camera and Human

    cs.CV 2025-09 conditional novelty 5.0 of 10

    A unified camera-and-human motion model with analytical heading decomposition and soft camera-trajectory integration improves global human trajectory reconstruction on RICH, EMDB, and 3DPW benchmarks.

Reference graph

Works this paper leans on

58 extracted references · 53 canonical work pages · cited by 2 Pith papers

  1. [1]

    https://www.flickr.com, 2024

    Flickr. https://www.flickr.com, 2024

    work page 2024

  2. [2]

    https://www.pexels.com/, 2024

    Pexels. https://www.pexels.com/, 2024

    2024

  3. [3]

    Objectron: A large scale dataset of object-centric videos in the wild with pose annota- tions

    Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jian- ing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annota- tions. In Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 7822–7831, 2021

    work page 2021

  4. [4]

    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. In Conference on Computer Vision and Pattern Recognition (CVPR), 2014

    work page 2014

  5. [5]

    Google street view: Capturing the world at street level

    Dragomir Anguelov, Carole Dulong, Daniel Filip, Christian Frueh, St ´ephane Lafon, Richard Lyon, Abhijit Ogale, Luc Vincent, and Josh Weaver. Google street view: Capturing the world at street level. pages 32–38, 2010

    work page 2010

  6. [6]

    Black, Priyanka Patel, Joachim Tesch, and Jin- long Yang

    Michael J. Black, Priyanka Patel, Joachim Tesch, and Jin- long Yang. BEDLAM: A synthetic dataset of bodies ex- hibiting detailed lifelike animated motion. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 8726–8737, 2023

    work page 2023

  7. [7]

    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. In European Conference on Computer Vision (ECCV), 2016

    work page 2016

  8. [8]

    Vasileios Choutas, Georgios Pavlakos, Timo Bolkart, Dim- itrios Tzionas, and Michael J. Black. Monocular expres- sive body regression through body-driven attention. InEuro- pean Conference on Computer Vision (ECCV), pages 20–40, 2020

    work page 2020

Show all 58 references
  1. [9]

    Huang, Siyu Tang, Dimitrios Tzionas, and Michael J

    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. In IEEE/CVF Conf. on Computer Vision and Pattern Recog- nition (CVPR), pages 2718–2728, 2022

    2022

  2. [10]

    ImageNet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition (CVPR), 2009

    2009

  3. [11]

    An image is worth 16x16 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 16x16 words: Transformers for image recognition ...

    2021

  4. [12]

    Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, Yao Feng, and Michael J. Black. TokenHMR: Advancing human mesh re- covery with a tokenized pose representation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024

    2024

  5. [13]

    Yao Feng, Vasileios Choutas, Timo Bolkart, Dimitrios Tzionas, and Michael J. Black. Collaborative regression of expressive bodies using moderation. In International Con- ference on 3D Vision (3DV), 2021

    2021

  6. [14]

    Vision meets robotics: The KITTI dataset

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The KITTI dataset. Inter- national Journal of Robotics Research , pages 1231 – 1237, 2013

    2013

  7. [15]

    Humans in 4D: Reconstructing and tracking humans with transformers

    Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa*, and Jitendra Malik*. Humans in 4D: Reconstructing and tracking humans with transformers. In International Conference on Computer Vision (ICCV), 2023

    2023

  8. [16]

    Ava: A video dataset of spatio-temporally localized atomic visual actions

    Chunhui Gu, Chen Sun, David A Ross, Carl V ondrick, Car- oline Pantofaru, Yeqing Li, Sudheendra Vijayanarasimhan, George Toderici, Susanna Ricco, Rahul Sukthankar, et al. Ava: A video dataset of spatio-temporally localized atomic visual actions. In Conference on Computer Visio...

    2018

  9. [17]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016

    2016

  10. [18]

    Huang, Hongwei Yi, Markus H ¨oschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J

    Chun-Hao P. Huang, Hongwei Yi, Markus H ¨oschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In Conference on Computer Vision and Pattern Recognition (CVPR), 2022

    2022

  11. [19]

    Human3.6M: Large scale datasets and predic- tive methods for 3D human sensing in natural environments

    Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6M: Large scale datasets and predic- tive methods for 3D human sensing in natural environments. Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 36(7):1325–1339, 2014

    2014

  12. [20]

    Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F. Fouhey. Perspective fields for single image camera calibration. In Conference on Computer Vision and Pattern Recognition (CVPR), 2023

    2023

  13. [21]

    Learning effective hu- man pose estimation from inaccurate annotation

    Sam Johnson and Mark Everingham. Learning effective hu- man pose estimation from inaccurate annotation. In Confer- ence on Computer Vision and Pattern Recognition (CVPR) , 2011

    2011

  14. [22]

    Ex- emplar fine-tuning for 3d human pose fitting towards in-the- wild 3d human pose estimation

    Hanbyul Joo, Natalia Neverova, and Andrea Vedaldi. Ex- emplar fine-tuning for 3d human pose fitting towards in-the- wild 3d human pose estimation. In International Conference on 3D Vision (3DV), 2020

    2020

  15. [23]

    Black, David W

    Angjoo Kanazawa, Michael J. Black, David W. Jacobs, and Jitendra Malik. End-to-end recovery of human shape and pose. In Conference on Computer Vision and Pattern Recog- nition (CVPR), pages 7122–7131, 2018

    2018

  16. [24]

    Learning 3d human dynamics from video

    Angjoo Kanazawa, Jason Y Zhang, Panna Felsen, and Ji- tendra Malik. Learning 3d human dynamics from video. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 5614–5623, 2019

    2019

  17. [25]

    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. In International Conference on Computer Vision (ICCV), 2023

    2023

  18. [26]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015

    2015

  19. [27]

    Beyond weak perspective for 9 monocular 3d human pose estimation

    Imry Kissos, Lior Fritz, Matan Goldman, Omer Meir, Ed- uard Oks, and Mark Kliger. Beyond weak perspective for 9 monocular 3d human pose estimation. In European Confer- ence on Computer Vision Workshops (ECCV-W), 2020

    2020

  20. [28]

    Huang, Joachim Tesch, Lea M ¨uller, Otmar Hilliges, and Michael J

    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 International Conference on Computer Vision (ICCV), 2021

    2021

  21. [29]

    Huang, Joachim Tesch, Lea M ¨uller, Otmar Hilliges, and Michael J

    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 In- ternational Conference on Computer Vision (ICCV) . IEEE, 2021

    2021

  22. [30]

    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. In International Conference on Computer Vision (ICCV), pages 2252–2261, 2019

    2019

  23. [31]

    CTRL-C: camera cali- bration transformer with line-classification

    Jinwoo Lee, Hyunsung Go, Hyunjoon Lee, Sunghyun Cho, Min-Hyuk Sung, and Junho Kim. CTRL-C: camera cali- bration transformer with line-classification. In International Conference on Computer Vision (ICCV), 2021

    2021

  24. [32]

    Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation

    Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 3383–3393, 2021

    2021

  25. [33]

    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. In Euro- pean Conference on Computer Vision, 2022

    2022

  26. [34]

    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 European Conference on Computer Vision (ECCV) , pages 740–755, 2014

    2014

  27. [35]

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multi- person linear model. Transactions on Graphics (TOG) , 34 (6):248:1–248:16, 2015

    2015

  28. [36]

    Monocular 3D human pose estimation in the wild using improved CNN supervision

    Dushyant Mehta, Helge Rhodin, Dan Casas, Pascal Fua, Oleksandr Sotnychenko, Weipeng Xu, and Christian Theobalt. Monocular 3D human pose estimation in the wild using improved CNN supervision. 2017

    2017

  29. [37]

    Priyanka Patel, Chun-Hao Paul Huang, Joachim Tesch, David Hoffmann, Shashank Tripathi, and Michael J. Black. AGORA: Avatars in geography optimized for regression analysis. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 13468–13478, 2021

    2021

  30. [38]

    Generating 3d faces using convolutional mesh autoencoders

    Anurag Ranjan, Timo Bolkart, Soubhik Sanyal, and Michael J Black. Generating 3d faces using convolutional mesh autoencoders. In Proceedings of the European confer- ence on computer vision (ECCV), pages 704–720, 2018

    2018

  31. [39]

    Syn- thetic training for accurate 3D human pose and shape esti- mation in the wild

    Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Syn- thetic training for accurate 3D human pose and shape esti- mation in the wild. In British Machine Vision Conference (BMVC), 2020

    2020

  32. [40]

    Hi- erarchical kinematic probability distributions for 3D human shape and pose estimation from images in the wild

    Akash Sengupta, Ignas Budvytis, and Roberto Cipolla. Hi- erarchical kinematic probability distributions for 3D human shape and pose estimation from images in the wild. In In- ternational Conference on Computer Vision (ICCV) , pages 11219–11229, 2021

    2021

  33. [41]

    Soyong Shin, Juyong Kim, Eni Halilaj, and Michael J. Black. Wham: Reconstructing world-grounded humans with accu- rate 3d motion. In Conference on Computer Vision and Pat- tern Recognition (CVPR), 2024

    2024

  34. [42]

    Indoor segmentation and support inference from rgbd images

    Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision (ECCV), pages 746–760, 2012

    2012

  35. [43]

    Deep high-resolution representation learning for human pose es- timation

    Ke Sun, Bin Xiao, Dong Liu, and Jingdong Wang. Deep high-resolution representation learning for human pose es- timation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  36. [44]

    Recovering accurate 3D human pose in the wild using IMUs and a moving camera

    Timo von Marcard, Roberto Henschel, Michael Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3D human pose in the wild using IMUs and a moving camera. In European Conference on Computer Vision (ECCV), pages 614–631, 2018

    2018

  37. [45]

    Refit: Recurrent fitting network for 3d human recovery

    Yufu Wang and Kostas Daniilidis. Refit: Recurrent fitting network for 3d human recovery. InInternational Conference on Computer Vision (ICCV), 2023

    2023

  38. [46]

    Ai challenger: A large-scale dataset for going deeper in image understanding

    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

    2017 arXiv

  39. [47]

    Detectron2

    Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2. https://github. com/facebookresearch/detectron2, 2019

    2019

  40. [48]

    Ehinger, Aude Oliva, and Anto- nio Torralba

    Jianxiong Xiao, Krista A. Ehinger, Aude Oliva, and Anto- nio Torralba. Recognizing scene viewpoint using panoramic place representation. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, 2012

    2012

  41. [49]

    SUN3D: A database of big spaces reconstructed using sfm and object labels

    Jianxiong Xiao, Andrew Owens, and Antonio Torralba. SUN3D: A database of big spaces reconstructed using sfm and object labels. In International Conference on Computer Vision (ICCV), pages 1625–1632, 2013

    2013

  42. [50]

    ViTPose: Simple vision transformer baselines for human pose estimation

    Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. ViTPose: Simple vision transformer baselines for human pose estimation. In Advances in Neural Information Pro- cessing Systems, 2022

    2022

  43. [51]

    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. In Conference on Computer Vision and Pat- tern Recognition (CVPR), 2021

    2021

  44. [52]

    On the continuity of rotation representations in neural networks

    Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5745–5753, 2019

    2019

  45. [53]

    Tame a wild camera: In-the-wild monocular camera calibra- tion

    Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibra- tion. Advances in Neural Information Processing Systems , 36, 2024. 10 CameraHMR: Aligning People with Perspective Supplementary Material Figure 5. Focal Length di...

    2024

  46. [54]

    The distribution shows notable peaks corresponding to the focal lengths of lenses most fre- quently used in photography, e.g

    Focal length distribution We plot the distribution of focal lengths used in training Hu- manFoV model in Figure 5. The distribution shows notable peaks corresponding to the focal lengths of lenses most fre- quently used in photography, e.g. 24, 28, 35, 50, 85, 105, 135, 200, 300,

  47. [55]

    Losses We use several loss functions to ensure accurate 3D human pose and shape estimation

    CameraHMR 8.1. Losses We use several loss functions to ensure accurate 3D human pose and shape estimation. We minimize the L2 norm dis- tance between the ground truth 3D joint locations ˆJ3d ∈ R44×3 and the predicted 3D joint locations J3d ∈ R44×3 centered around pelvis joint ...

  48. [56]

    As describe in Eq

    CamSMPLify Here we provide more details about the optimization proce- dure used for generating our pseudo ground truth data for 4DHumans dataset. As describe in Eq. 5 from the main paper, we minimize the energy term E(β, θ, tfull) by opti- mizing for SMPL shape β and pose θ as...

  49. [57]

    To address this, we utilize the SSP-3D [39] dataset, which includes a broad spectrum of body shapes

    Shape Evaluation Most HPS evaluation benchmarks primarily represent aver- age body shapes and offer limited shape diversity, which restricts their effectiveness in assessing improvements in shape accuracy. To address this, we utilize the SSP-3D [39] dataset, which includes a b...

  50. [58]

    6 and Fig

    More Qualitative Results In Fig. 6 and Fig. 7, we present qualitative results of Cam- eraHMR applied to images downloaded from Pexels [2]. The results for multi-person images are obtained by first generating the bounding box for each person using De- tectron2 [47] on the full ...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.