Pith. sign in

REVIEW 4 major objections 6 minor 19 references

Biomechanical Reconstruction with Confidence Intervals from Multiview Markerless Motion Capture

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read By fitting a Gaussian variational posterior over the joint-angle trajectory end-to-end through a differentiable biomechanical model, this paper turns multiview markerless motion capture into a source of calibrated, per-trial confidence…

desk verdict A clean variational extension of the authors' deterministic MMMC pipeline that gives useful per-trial uncertainty, but the well-calibrated confidence interval claim is only demonstrated against the same keypoints used for fitting, not against true anatomical ground truth. read the letter →

arxiv 2502.06486 v1 pith:OZKZOJ5B submitted 2025-02-10 cs.CV

R. James Cotton , Fabian Sinz This is my paper · ORCID
classification cs.CV
keywords markerlessmotioncapturevariationalinferenceconfidenceintervalsbiomechanicalmodeluncertaintycalibrationmultiviewposeestimationjointanglesexpectederror
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 asks whether a single markerless motion-capture trial—recorded with an arbitrary camera configuration, on a patient who may move abnormally—can report not just a kinematic estimate but a trustworthy uncertainty for that estimate. It claims the answer is yes: a Gaussian variational posterior over the joint-angle trajectory, fitted end-to-end through a differentiable biomechanical model, produces confidence intervals for each joint at each moment, with median expected calibration error of 0.07 at a one-pixel clip and 0.0005 at a two-pixel clip. This matters because clinical practice and large-scale movement studies need to know whether a particular measurement can be trusted, whereas traditional validation only describes average performance. The paper reports spatial intervals typically within 10–15 mm and joint-angle intervals of only a few degrees, widening for distal joints.

What carries the argument

The load-bearing object is a Gaussian variational posterior $q_\phi(\theta_t) = \mathcal{N}(\mu_\phi(t), \Sigma_\phi(t))$ over the 40-dimensional joint-angle vector, with a low-rank covariance parameterized by an MLP that maps time to $(\mu, u)$. Reparameterized samples are pushed through a differentiable biomechanical forward-kinematics model and camera projections, then scored against a likelihood over Euclidean reprojection error whose width $\sigma_\psi(s) = \mathrm{softplus}(\psi_0) + \mathrm{softplus}(\psi_1) s + \mathrm{softplus}(\psi_2) s^2$ is a learned second-order polynomial of the detector's confidence score. Maximizing the ELBO trades entropy against keypoint evidence, so inconsistencies across views or time widen the intervals. Calibration is checked with an expected calibration error computed on the highest-confidence keypoints, with the learned likelihood width clipped to isolate the posterior's contribution.

What would settle it

Run synchronized biplanar fluoroscopy or marker-based motion capture on the same trials and compare the predicted 95% posterior intervals for bone positions and joint angles against that near-ground-truth data; the central claim fails if actual coverage is substantially below 95% on trials where the high-confidence-keypoint expected calibration error looks good, or if the expected calibration error computed over all keypoints rather than the best 5% degrades sharply.

Watch

Extended reading notes

Core claim

The central claim is that posterior uncertainty over pose can be inferred from the same multiview keypoint detections used for reconstruction, without ground-truth markers. The method represents the trajectory as an implicit function of time that outputs both a mean pose and a low-rank covariance, optimizes the evidence lower bound by sampling poses, projecting them through forward kinematics and camera calibration, and evaluating them against a learned likelihood whose width grows with the keypoint detector's reported noise. The resulting per-trial, per-joint confidence intervals are well calibrated as measured by expected calibration error on high-confidence keypoints, and the posterior captures correlations such as hip–pelvis coupling. The paper also shows that injected keypoint noise is recovered in the learned likelihood width and that removing cameras widens the intervals, both as expected.

Load-bearing premise

The calibration claim rests on the assumption that the 5th percentile of most confidently detected keypoints lies near the true keypoint locations and that the learned likelihood, clipped at a fixed pixel width, accounts for all remaining detection noise; if keypoint detections are systematically biased, for example by viewpoint, the reported calibration error will not reveal the bias and the posterior can be overconfident.

Editorial extensions

If this is right

  • Clinicians can decide per measurement whether to trust a joint angle at a given moment, rather than relying only on population-level validation.
  • Segments of a trial with occlusions, few cameras, or poor geometry will carry visibly wider intervals, allowing automatic exclusion of unreliable data.
  • The learned correlation structure, such as hip–pelvis coupling, quantifies how error in one joint angle propagates to another.
  • Uncertainty estimates become available for a novel individual and camera configuration without any ground-truth instrumentation.
  • Spatial intervals generally staying within 10–15 mm, and proximal joint angles within a few degrees, means the posterior is informative enough to be clinically useful.

Reading between the lines

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

  • Editorial extension: if the intervals are as well calibrated as reported, they could serve as observation-noise models for downstream gait classification or biomechanical simulation, a use the paper does not implement.
  • Editorial extension: because the reported expected calibration error depends on clipping the learned likelihood width, a stricter check would be marker-based validation on the same trials; the paper identifies this as future work.
  • Editorial extension: the same variational machinery could be applied per camera or per keypoint to expose which views drive the uncertainty, potentially guiding active camera placement.
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

4 major / 6 minor

Summary. The paper extends a deterministic end-to-end markerless motion capture pipeline to a probabilistic variational framework. An implicit neural representation maps time to the mean and a low-rank covariance of a Gaussian posterior over joint angle trajectories; the posterior is optimized by maximizing an ELBO that includes a learned likelihood over detected keypoints, whose width is a second-order polynomial of the detector's confidence score. The authors report per-joint confidence intervals for joint angles and virtual marker positions, and validate them using an expected calibration error (ECE) computed in image space against high-confidence detected keypoints, with a clipping parameter on the learned likelihood width. The paper also reports population statistics from 2957 trials, including spatial errors, joint angle uncertainties, and ECE values.

Significance. If the central claim were established, the method would be a practically useful contribution to clinical movement analysis, since it addresses the need for per-trial, per-joint confidence intervals rather than population-level averages. The paper also demonstrates a technically interesting extension of an existing differentiable biomechanics pipeline to a probabilistic setting, and it explicitly attempts to learn the keypoint likelihood function. However, the reported calibration evidence is not sufficient to support the headline claim: the ECE is computed on the same keypoints used to fit the posterior and the likelihood, and it depends on an adjustable clipping parameter. The paper's own Discussion acknowledges that viewpoint-dependent biases can produce overconfident estimates, which the current validation cannot detect. The potential significance is high, but the current manuscript does not yet validate the method against an independent ground truth.

major comments (4)
  1. [§3.5.3 and Appendix] The ECE is computed on the same detected keypoints that are used to fit the variational posterior and the likelihood parameters. Since φ and ψ are optimized to maximize the ELBO on these exact keypoints, the near-zero ECE at a 2-pixel clip only demonstrates that the predictive distribution covers its own training data. The appendix's pseudo-ground-truth selection (5th percentile of highest-confidence keypoints) does not break this circularity because those keypoints are also part of the fitting set. The central claim of well-calibrated per-trial confidence intervals requires out-of-sample validation, for example against marker-based motion capture or biplanar fluoroscopy, or at least a held-out set of trials, cameras, or keypoints.
  2. [§2.2 and Appendix] The reported ECE values depend strongly on the clip parameter σ_ECE: the median ECE drops from 0.33 at σ_ECE=0 to 0.07 at 1 pixel and 5e-4 at 2 pixels. Since the paper states it monitors ECE for several σ_ECE values from 0 to 5 pixels, reporting the 2-pixel clip as evidence of calibration is equivalent to selecting the threshold that yields the desired result. The ECE at σ_ECE=0 is the more direct measure of whether the posterior alone covers the keypoints, and its value of 0.33 indicates substantial miscalibration. The manuscript should either justify a single pre-registered clip level or treat the clip as a hyperparameter whose effect is analyzed, not as part of the validation result.
  3. [§3.5.3 and Appendix] The ECE computation is inconsistent with the model's likelihood. The likelihood is an exponential distribution over the radial error, but the ECE is computed by taking the Gaussian pushforward of the posterior, adding the likelihood variance, and then using a Rayleigh CDF for the radial error. The resulting distribution is not the actual predictive distribution of the model, which would be a convolution of a Gaussian and an exponential-in-radius distribution. Therefore, the reported ECE does not measure the calibration of the model's true posterior predictive, and the numerical values cannot be directly interpreted as confidence-interval coverage.
  4. [§3.4.3 and Discussion] The lesion experiment injects zero-mean noise and confirms that the likelihood width adapts to the added variance, but it does not test the failure mode identified in the Discussion: viewpoint-dependent errors in keypoint locations can produce biased results while being overconfident. A concrete experiment with a constant or viewpoint-dependent offset on a subset of keypoints would show whether the posterior confidence intervals widen appropriately when the detector is biased. Without such a test, the claim that the method provides confidence intervals reflecting true accuracy remains conditional on the detector being unbiased conditional on its confidence score.
minor comments (6)
  1. [§3.5.3] Typo: 'vartional' should be 'variational'.
  2. [§3.5.4] Typo: 'lumber' should be 'lumbar' in the correlation-structure figure caption and text.
  3. [§3.5.3] Typo: 'ours distributions' should be 'our distributions'.
  4. [§2.1.5] The sentence 'The probability functions was optimized using the reparameterization trick' contains a subject-verb agreement error; it should be 'The probability function was optimized'.
  5. [§2.1.4] The paper says 'the score measures predicted noise and increases for less accurate keypoints', but many keypoint detectors output a confidence score that increases with accuracy. Please clarify the convention explicitly, since the sign of the score dependence is central to the likelihood parameterization.
  6. [Figure 8] The bottom panel of Figure 8 uses different units (meters) than the top panel (mm), but the y-axis labels are not visually clear. Please label each panel with its units to avoid confusion.

Circularity Check

1 steps flagged · score 6.0 of 10

Calibration ECE is computed on the same detected keypoints used to fit the posterior and likelihood, so the near-zero ECE at the 2-pixel clip largely reflects in-sample coverage rather than independent verification of joint-angle confidence intervals.

  1. fitted input called prediction [Section 2.2 and Section 3.5.3; Appendix (ECE metric)]
    "Instead, we measure ECE with respect to the detected keypoints, although they are not ground truth. ... we approximate ground-truth keypoints using the 5th percentile of keypoints with the highest detection confidence. ... we clip the width of the likelihood function to a maximum value, σ = min(σψ(s),σ ECE) prior to adding it to the posterior variance and computing the ECE."

    The variational posterior q_ϕ and likelihood width σ_ψ are jointly optimized to maximize the ELBO on all detected keypoints, including the high-confidence subset later used as pseudo-ground truth. The ECE then evaluates predicted keypoint distributions against those same fitted keypoints, so a low ECE (especially with the 2-pixel likelihood floor) mainly confirms that the fitted distribution covers its own training observations. This does not establish that the 95% joint-angle confidence intervals contain the true anatomical angles; the paper's appendix itself acknowledges the failure mode where an overestimated likelihood masks an underestimated posterior. The external walkway comparison is the only independent calibration support.

full rationale

The variational derivation itself is self-contained: the posterior and likelihood are optimized end-to-end through a differentiable biomechanical model, and the lesion and camera-removal experiments provide independent evidence that uncertainty increases with injected noise and fewer views. However, the headline claim of well-calibrated per-joint confidence intervals rests on an ECE computed in image space against the same detected keypoints used for fitting. The high-confidence pseudo-ground-truth subset is part of the training keypoints, and the likelihood width is clipped and additively applied before computing the ECE, so the near-zero ECE at a 2-pixel clip is largely an in-sample coverage measure rather than external calibration against true pose. The discussion and appendix explicitly acknowledge the absence of ground-truth poses and the possibility of biased, overconfident results from viewpoint-dependent errors. This is a genuine circular-validation gap in the central claim, though not a derivation identical to its inputs; the method's response to synthetic noise and camera count supplies independent content. Score 6 reflects partial circularity in the calibration validation without dismissing the rest of the contribution.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

The method rests on a standard variational inference objective plus several modeling choices: a Gaussian posterior family, a polynomial likelihood width in the keypoint confidence score, a learned skeleton scale, a chosen covariance rank, and an ad hoc clipping rule for calibration. No new physical entities are introduced.

free parameters (5)
  • psi0, psi1, psi2 (likelihood width coefficients) = not reported (learned during optimization)
    Width of the exponential keypoint error distribution is sigma_psi(s) = softplus(psi0) + softplus(psi1)s + softplus(psi2)s^2; fitted jointly with the posterior and central to the reported calibration.
  • beta (body scale and marker offsets) = not reported (learned per participant)
    Personalized skeleton scale and virtual marker offsets optimized in a bilevel loop; these affect all forward kinematic projections and therefore all reported uncertainty estimates.
  • R (low-rank covariance rank) = 20
    Selected from an ELBO and ECE sweep; larger ranks slightly increase entropy and reduce ECE_0, with a plateau near rank 20.
  • sigma_ECE clipping level = 0, 1, and 2 pixels used for reported ECE
    Ad hoc clipping of the learned likelihood width before computing calibration, intended to prevent the likelihood from absorbing too much uncertainty; clipping to 2 pixels makes ECE nearly zero.
  • Keypoint likelihood distribution family = Exponential
    Chosen among Half Normal, Exponential, and Half Cauchy based on ELBO; Half Normal diverged. This is a modeling choice rather than a numeric parameter.
assumptions (7)
  • standard math Bayes' rule and the ELBO decomposition are valid for the posed inference problem
    Section 2.1.2 derives the evidence lower bound and uses the reparameterization trick; this is standard variational inference.
  • domain assumption A Gaussian posterior with low-rank covariance sufficiently represents true pose uncertainty
    Section 2.1.3; the variational family may be misspecified for multimodal or strongly non-Gaussian pose posteriors, especially under occlusion.
  • domain assumption The keypoint confidence score is a monotone proxy for keypoint error and the likelihood width is a second-order polynomial of the score
    Section 2.1.4; the detector's uncalibrated confidence scores are assumed usable as parameters of the likelihood function.
  • domain assumption The highest-confidence 5% of detected keypoints are near ground truth
    Appendix 'Metric: Expected Calibration Error'; these pseudo-ground-truth keypoints are the basis for all ECE claims, so any bias in them invalidates the calibration assessment.
  • domain assumption First-order Taylor expansion of the pushforward through kinematics and camera projection is accurate
    Appendix 'Taylor approximation of q_yhatc(yhatc)'; linearization may fail when the posterior covariance is large, e.g., during occlusions.
  • domain assumption Uniform prior over joint limits does not bias clinical populations
    Section 2.1.2; chosen to avoid biasing toward able-bodied poses, but it is a weak and uninformative prior.
  • ad hoc to paper The sigma clipping rule sigma = min(sigma_psi, sigma_ECE) is a valid correction for calibration assessment
    Appendix; introduced to stop the learned likelihood from absorbing too much uncertainty, but it changes the statistical meaning of the ECE and is not derived from first principles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Biomechanical Reconstruction with Confidence Intervals from Multiview Markerless Motion Capture." pith.science (2026). https://pith.science/paper/OZKZOJ5B

@misc{pith2026250206486,
  author       = {Pith},
  title        = {Pith review of: Biomechanical Reconstruction with Confidence Intervals from Multiview Markerless Motion Capture},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OZKZOJ5B}},
  note         = {Machine review of arXiv:2502.06486}
}
read the original abstract

Advances in multiview markerless motion capture (MMMC) promise high-quality movement analysis for clinical practice and research. While prior validation studies show MMMC performs well on average, they do not provide what is needed in clinical practice or for large-scale utilization of MMMC -- confidence intervals over specific kinematic estimates from a specific individual analyzed using a possibly unique camera configuration. We extend our previous work using an implicit representation of trajectories optimized end-to-end through a differentiable biomechanical model to learn the posterior probability distribution over pose given all the detected keypoints. This posterior probability is learned through a variational approximation and estimates confidence intervals for individual joints at each moment in a trial, showing confidence intervals generally within 10-15 mm of spatial error for virtual marker locations, consistent with our prior validation studies. Confidence intervals over joint angles are typically only a few degrees and widen for more distal joints. The posterior also models the correlation structure over joint angles, such as correlations between hip and pelvis angles. The confidence intervals estimated through this method allow us to identify times and trials where kinematic uncertainty is high.

Figures

Figures reproduced from arXiv: 2502.06486 by the authors.

Figure 1
Figure 1. Overview of our approach. Samples from the variational posterior distribution, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Influence of distribution rank. The left panel shows the distribution entropy relative to the 0-rank [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Influence of artificially injected keypoint noise. The left panel shows the distribution entropy [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: The influence of removing cameras. The left panel shows the distribution entropy relative to the [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The top row shows the 50th percentile spatial confidence interval for the body parts. The second [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: The top row shows the 50th percentile joint angle error interval for the joints. The second row [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: The left panel shows the average correlation structure focused on joints in the lower body. The [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: Each subpanel shows the joint angle and position estimates with their 95th percentile confidence [PITH_FULL_IMAGE:figures/full_fig_p011_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 15 canonical work pages

  1. [1]

    James Cotton

    R. James Cotton. Differentiable Biomechanics Unlocks Opportunities for Markerless Motion Capture , 2 2024. [Online; accessed 2024-02-28]

    work page 2024

  2. [2]

    Mujoco: A Physics Engine for Model - Based Control

    Emanuel Todorov, Tom Erez, and Yuval Tassa. Mujoco: A Physics Engine for Model - Based Control . In 2012 IEEE / RSJ International Conference on Intelligent Robots and Systems , pages 5026--5033, 10 2012. doi:10.1109/IROS.2012.6386109. [Online; accessed 2023-11-06]

    arXiv 2012

  3. [3]

    Myosuite -- A Contact - Rich Simulation Suite for Musculoskeletal Motor Control , 5 2022

    Vittorio Caggiano, Huawei Wang, Guillaume Durandau, Massimo Sartori, and Vikash Kumar. Myosuite -- A Contact - Rich Simulation Suite for Musculoskeletal Motor Control , 5 2022. [Online; accessed 2023-12-28]

    work page 2022

  4. [4]

    Multiple View Geometry in Computer Vision

    Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision . Cambridge University Press, 2003. ISBN 978-0-521-54051-3

    work page 2003

  5. [5]

    Equinox: Neural Networks in JAX via Callable PyTrees and Filtered Transformations

    Patrick Kidger and Cristian Garcia. Equinox: Neural Networks in JAX via Callable PyTrees and Filtered Transformations . Differentiable Programming workshop at Neural Information Processing Systems 2021, 10 2021. [Online; accessed 2021-11-02]

    work page 2021

  6. [6]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto- Encoding Variational Bayes . In 2nd International Conference on Learning Representations , ICLR 2014 - Conference Track Proceedings . International Conference on Learning Representations, ICLR, dec 20 2014

    work page 2014

  7. [7]

    Decoupled Weight Decay Regularization , 1 2019

    Ilya Loshchilov and Frank Hutter. Decoupled Weight Decay Regularization , 1 2019. [Online; accessed 2023-07-02]

    work page 2019

  8. [8]

    The DeepMind JAX Ecosystem , 2020

    DeepMind , Igor Babuschkin, Kate Baumli, Alison Bell, Surya Bhupatiraju, Jake Bruce, Peter Buchlovsky, David Budden, Trevor Cai, Aidan Clark, Ivo Danihelka, Antoine Dedieu, Claudio Fantacci, Jonathan Godwin, Chris Jones, Ross Hemsley, Tom Hennigan, Matteo Hessel, Shaobo Hou, Steven Kapturowski, Thomas Keck, Iurii Kemaev, Michael King, Markus Kunesch, Lena...

    work page 2020

Show all 19 references
  1. [9]

    Dillon, Ian Langmore, Dustin Tran, Eugene Brevdo, Srinivas Vasudevan, Dave Moore, Brian Patton, Alex Alemi, Matt Hoffman, and Rif A

    Joshua V. Dillon, Ian Langmore, Dustin Tran, Eugene Brevdo, Srinivas Vasudevan, Dave Moore, Brian Patton, Alex Alemi, Matt Hoffman, and Rif A. Saurous. Tensorflow Distributions , 11 2017. [Online; accessed 2024-03-05]

    2017

  2. [10]

    Hicks, Steve Collins, Scott L

    Keenon Werling, Michael Raitor, Jon Stingel, Jennifer L. Hicks, Steve Collins, Scott L. Delp, and C. Karen Liu. Rapid Bilevel Optimization to Concurrently Solve Musculoskeletal Scaling , Marker Registration , and Inverse Kinematic Problems for Human Motion Reconstruction , 8 2...

    2022

  3. [11]

    Distribution Calibration for Regression

    Hao Song, Tom Diethe, Meelis Kull, and Peter Flach. Distribution Calibration for Regression . In Proceedings of the 36th International Conference on Machine Learning , pages 5897--5906. PMLR, 5 2019. [Online; accessed 2024-03-03]

    2019

  4. [12]

    Pierzchlewicz, Mohammad Bashiri, R

    Pawe A. Pierzchlewicz, Mohammad Bashiri, R. James Cotton, and Fabian H. Sinz. Optimizing MPJPE Promotes Miscalibration in Multi - Hypothesis Human Pose Lifting . In ICLR 2023 Affinity Workshop , 4 2023. [Online; accessed 2023-05-30]

    2023

  5. [13]

    Fast and Robust Multi - Person 3d Pose Estimation and Tracking from Multiple Views

    Junting Dong, Qi Fang, Wen Jiang, Yurou Yang, Hujun Bao, and Xiaowei Zhou. Fast and Robust Multi - Person 3d Pose Estimation and Tracking from Multiple Views . In T- PAMI , 2021

    2021

  6. [14]

    Learning 3d Human Pose Estimation from Dozens of Datasets Using a Geometry - Aware Autoencoder to Bridge Between Skeleton Formats

    Istv \' a n S \' a r \' a ndi, Alexander Hermans, and Bastian Leibe. Learning 3d Human Pose Estimation from Dozens of Datasets Using a Geometry - Aware Autoencoder to Bridge Between Skeleton Formats . In IEEE/ CVF Winter Conference on Applications of Computer Vision ( WACV ) ....

    2023 doi arXiv

  7. [15]

    Saeed Ghorbani, Kimia Mahdaviani, Anne Thaler, Konrad Kording, Douglas James Cook, Gunnar Blohm, and Nikolaus F. Troje. Movi: A Large Multi - Purpose Human Motion and Video Dataset . PLOS ONE, 16 0 (6): 0 e0253157, 6 2021. ISSN 1932-6203. doi:10.1371/journal.pone.0253157. [Onl...

    2021 doi

  8. [16]

    James Cotton, Allison DeLillo, Anthony Cimorelli, Kunal Shah, J

    R. James Cotton, Allison DeLillo, Anthony Cimorelli, Kunal Shah, J. D. Peiffer, Shawana Anarwala, Kayan Abdou, and Tasos Karakostas. Optimizing Trajectories and Inverse Kinematics for Biomechanical Analysis of Markerless Motion Capture Data . In IEEE International Consortium f...

    2023 doi arXiv

  9. [17]

    James Cotton, Anthony Cimorelli, Kunal Shah, Shawana Anarwala, Scott Uhlrich, and Tasos Karakostas

    R. James Cotton, Anthony Cimorelli, Kunal Shah, Shawana Anarwala, Scott Uhlrich, and Tasos Karakostas. Improved Trajectory Reconstruction for Markerless Pose Estimation . In 45th Annual International Conference of the IEEE Engineering in Medicine and Biology Society , Sydney, ...

    2023

  10. [18]

    Experiment Tracking with Weights and Biases , 2020

    Lukas Biewald. Experiment Tracking with Weights and Biases , 2020

    2020

  11. [19]

    Peiffer, Kunal Shah, Lee E Miller, and R

    Pouyan Firouzabadi, Wendy Murray, Anton R Sobinov, J.D. Peiffer, Kunal Shah, Lee E Miller, and R. James Cotton. Biomechanical Arm and Hand Tracking with Multiview Markerless Motion Capture . In 2024 10th IEEE RAS / EMBS International Conference for Biomedical Robotics and Biom...

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.