Pith. sign in

REVIEW 3 major objections 6 minor 69 references

Posterior shape models revisited: Improving 3D reconstructions from partial data using target specific models

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Projecting a pre-trained shape model onto the observed domain recovers the target-aligned model—exactly for translations, approximately for rotations—and improves partial-shape reconstruction accuracy and variance.

desk verdict Solid and practical: exact translation projection, honest rotation approximation, and a plug-and-play fix for prebuilt shape models; main gap is the absence of error bars and a real-target pose measurement. read the letter →

arxiv 2509.05776 v1 pith:LNRPNCNS submitted 2025-09-06 cs.CV

Jonathan Aellen , Florian Burkhardt , Thomas Vetter , Marcel Lüthi This is my paper
classification cs.CV
keywords shapereconstructionpointdistributionmodelsGaussianprocessmorphablegeneralizedProcrustesanalysisposealignmentposteriorpartialdatauncertaintycalibration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the pose alignment of training shapes relative to the partial target is a decisive, often overlooked factor in statistical shape reconstruction: models built on globally aligned data are biased and overconfident when only part of a shape is observed. It introduces a projection operator that realigns any existing point distribution model to the observed domain X in a single preprocessing step, with no access to the original training data. For translations the projection is exact; for small rotations it is a good linear approximation. The paper shows that the projected model matches a model retrained on target-aligned data across femur, scaphoid, and skull experiments, using three different inference methods, and that the simpler inference landscape also speeds convergence.

What carries the argument

The projection operator P^X, the complement of the translation-averaging operator T^X over the observed domain X, applied to the mean and covariance of a Gaussian process morphable model (GPMM) as GP(P^X µ^Z, P^X k^Z P^{X T}). Its kernel is spanned by normalized translation basis vectors and, for the rotation extension, discretized derivatives of rotations about X's center; its image is the space of shapes aligned on X. A discretized implementation projects the Karhunen–Loève basis vectors one by one, followed by a re-orthogonalization via SVD, at O(N r²) cost.

What would settle it

Run the paper's hinge-parameter experiment (D.1) at φ=0.3 with the true distribution having arm angles near ±90°: the projected model should systematically misestimate the observed arm's length and the predicted arm's variance relative to the target-aligned model—quantifying exactly the angle beyond which the projection no longer approximates a target-specific prior.

Watch

Extended reading notes

Core claim

The central claim is that target-specific priors should always be used for partial-shape reconstruction, and that they are always available: the target-specific model GP(µ^X, k^X) can be obtained from any pre-existing model GP(µ^Z, k^Z) by the linear projection P^X = I − T^X, where T^X averages a deformation field over the observed domain X. Because shapes in different alignments differ only by rigid motions, the projection's kernel—constant translations, plus, by extension, rotation derivatives around the center of X—is exactly what differentiates the alignments. The paper proves the equality µ^X = P^X µ^Z and k^X = P^X k^Z P^{X T} for the translation-only case, shows the extended kernel ap

Load-bearing premise

The load-bearing premise is that orientation normalization can be approximated by the linear derivative of rotation around the mean (small angles); the paper's own synthetic example (Appendix D.1) shows the approximation degrades and ultimately fails as rotations approach and exceed 90°, so real targets with large pose differences would break the claimed equivalence.

Editorial extensions

If this is right

  • Any existing prebuilt shape model—empirical or kernel-based, local or global—can be adapted to a specific partial target by a cheap preprocessing step, removing the need to retrain on the original data.
  • Reconstruction accuracy improves markedly when little of the target is observed (20–45% in the femur experiment), with mean squared error roughly halved compared to target-agnostic models.
  • Predicted variance becomes trustworthy: target-agnostic posteriors are overconfident, while the projected and target-specific models produce wide, realistic uncertainty on unobserved regions.
  • The inference problem becomes easier—pose and shape are disentangled—so methods like non-rigid ICP converge in fewer iterations and are less prone to local minima; even a 1.5-hour normalizing-flow run benefits.
  • The self-consistency check recovers the prior with the projected model, confirming the posterior is no longer biased—an indicator that should transfer to any Bayesian shape-reconstruction pipeline.

Reading between the lines

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

  • The same projection logic should apply to any linear latent-variable model with nuisance rigid parameters (e.g., morphable appearance or depth models), suggesting a general alignment-calibration layer that can be inserted without retraining.
  • The paper's small-angle caveat implies a practical boundary: if an initial rigid alignment exceeds roughly 15–20°, the projected model should be refreshed by iterating the estimate of X, or the reconstruction should be flagged as unreliable.
  • Because the projection is rank-preserving on the shape space and leaves Mahalanobis geometry intact, downstream shape statistics (sex/stature regression, classification) trained on the original model may inherit better-calibrated posteriors without modification.
  • The prior-recovery (self-consistency) test used here is a cheap diagnostic that could be adopted more widely to detect alignment mismatch in Bayesian models with nuisance pose.
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 / 6 minor

Summary. The paper revisits posterior shape models for reconstructing 3D shapes from partial observations. It argues that the rigid alignment of training data relative to a partial target strongly biases reconstructions and underestimates uncertainty when the model is not aligned on the observed domain. The authors formalize positional alignment as a projection operator and show that applying this operator to an existing Gaussian process morphable model's mean and covariance exactly yields the model that would be obtained by training on translation-aligned data (Eq. (11)). For rotational alignment, they extend the projection kernel with derivatives of rotations about the observed domain (Sec. 4.1, Eq. (13)), which they state is a small-angle approximation. Experiments on femur, scaphoid, and skull data compare target-agnostic, target-specific, and the proposed projected models under three inference schemes (nICP, Metropolis-Hastings, normalizing flows), reporting improved reconstruction error and more realistic predictive variance for the projected model.

Significance. If the central claim holds, the paper provides a practical, low-cost preprocessing step that adapts any existing linear shape model to a target-specific alignment without access to training data. The translation case is exact and cleanly derived, and the method is computationally efficient (O(N r^2)). The experimental coverage is broad: three inference methods, three anatomical datasets, and a public code repository. The paper also correctly identifies and demonstrates a real limitation of target-agnostic models. However, the rotational part of the method is an acknowledged approximation whose validity for real targets is not directly measured, and the experimental comparisons are reported as point estimates without variability or significance quantification. These issues concern the generalizability of the plug-and-play claim and the strength of the empirical conclusions, respectively.

major comments (3)
  1. [Sec. 4.1, Eq. (13)] The exact identity in Eq. (11) holds only for the translation-only alignment. The extension to the full GPA setting relies on approximating the nonlinear optimal-rotation operator by its first-order derivative, yet the manuscript does not prove that the projected covariance P^X k^Z P^{X^T} with the added rotation basis approximates the covariance of shapes aligned by iterated Procrustes. Section D.1 itself shows that the linear approximation degrades as the rotation variance grows, with clear failure at angles near 90 degrees. To support the plug-and-play claim, the authors should either derive the approximation error bound for the rotation case or measure the residual rotation distribution in the femur/scaphoid targets and show it is within the small-angle regime where the approximation is controlled.
  2. [Tables 1-4 and Figs. 3, 4, 11, 12] All quantitative results are reported as averages over leave-one-out folds without standard deviations, confidence intervals, or paired significance tests. The claim that target-specific and projected models are 'significantly' better than target-agnostic models is therefore not statistically substantiated. This matters because some differences between pP and pX are small (e.g., Table 1, 80% observed, nICP: 63 vs 63; MH: 23 vs 23) and because with normalizing flows the target-agnostic model nearly closes the gap in some columns (Table 1, NF, 60%: 47 vs 46; 80%: 25 vs 25). Please report per-fold variability and perform paired tests across folds.
  3. [Sec. 5.1 / Sec. B.2] The projected model pP is evaluated using an iteratively estimated X, while the target-specific pX uses the ground-truth X. This is a reasonable practical comparison, but the behavior of the X estimator is not analyzed: there is no discussion of initialization, convergence, or failure cases for X estimation. Since the exact translation identity assumes X is known, the robustness of the overall method for plug-and-play use depends on this estimator. Please clarify how X is estimated in each inference method and whether the ablation 'ours GT-X' (Table 2 and 3) is the only fair comparison for testing the projection itself.
minor comments (6)
  1. [Sec. 1, Figure 1] The caption says '±3 pc1' and 'pc1' but the panels are not all clearly labeled; consider adding explicit labels for prior, posterior, and target-agnostic/target-specific distributions directly in the panels.
  2. [Sec. 2] The related-work paragraph on diffeomorphic methods is very brief and somewhat disjointed from the rest of the paper. It may be clearer to focus on work that explicitly addresses partial shape reconstruction and pose alignment.
  3. [Sec. 4, Algorithm 1] The notation M_X and the line 'M_X ← M(1...3N_X, ::)' is unclear: the start/end indices for the observed domain X need a precise definition. Also, the normalization of the rotation basis vectors in Sec. 4.1 is not specified; please state the inner product used.
  4. [Sec. 5] The three inference methods differ greatly in computational cost and implementation. While this is intentional, a sentence summarizing which differences are due to prior alignment versus inference capacity would help avoid overinterpreting small numerical differences.
  5. [Appendix D.1] The text says 'For most biological shapes rotations of parts that exceed such angles are rare.' This is an empirical claim without citation or supporting experiment. Either add evidence or soften the statement to a conjecture.
  6. [General] There are several minor typos and grammatical issues, e.g., 'adaption' (Sec. 2), 'a powerful distribution' (Sec. B.3), 'the landmark aligned models ... it is preferable' (Sec. C). A careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: Eq. (11) is a linear-operator identity derivable from the paper's own alignment definition, and the rotation approximation is an explicitly tested modeling approximation rather than a fitted parameter.

full rationale

The central derivation in Sec. 4 is self-contained and non-circular. For the translation-only setting, P^X is defined as the complement of the translation operator T^X appearing in the GPA alignment objective Eq. (2), and Eq. (9) follows directly from that definition. Equation (11) is then an algebraic identity: because P^X annihilates constant deformation fields, one has P^X P^Z = P^X, so projecting the existing mean and covariance under P^X reproduces the empirical mean and covariance of the training shapes aligned on X. This is a mathematical consequence, not a fitted or renamed quantity. The rotation extension in Sec. 4.1 is introduced as an explicit first-order approximation, with its error regime analyzed in Sec. D.1; it is compared, not conflated, with independently constructed target-specific models built using full GPA, and it is not tuned to match reconstruction outcomes. The only target-dependent input is the observed domain X, which is part of the problem statement rather than a parameter fit to the ground-truth reconstruction. Self-citations to GPMM (Lüthi et al. 2017) and posterior shape models (Albrecht et al. 2013) provide background formalism; the projection proof does not depend on their correctness. The paper's own stated limitation that nonlinear orientation normalization requires a sufficiently accurate linear approximation is a scope condition, not evidence of circularity. Accordingly, no load-bearing step reduces by construction or by self-citation to its own inputs.

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

The method introduces no fitted parameters in the core derivation; the projection kernel is derived from the alignment objective, and the rotation derivative is a first-order approximation, not a fit to target outcomes. The listed domain assumptions are standard for PDM/GPMM-based reconstruction.

assumptions (4)
  • domain assumption The observed domain X is a subset of the reference domain Omega and is estimable from the target.
    The projection P^X is defined on the domain X observed on the target; its accuracy depends on the quality of the estimate of X (Sec 4 and best practices in Sec 6.1).
  • domain assumption Rotations of the observed domain between training data and target are small enough that the linear approximation via derivatives in Eq. (13) is valid.
    Sec 4.1 states the error scales with the sine of the rotation angle, and Sec D.1 shows the approximation breaks down for large angles.
  • domain assumption The shape distribution is Gaussian and can be represented by a linear GPMM basis.
    The method is restricted to linear models; the paper focuses on PDMs/GPMMs and acknowledges non-linear models are out of scope (Sec 7).
  • domain assumption Closest-point correspondence is a sufficient approximation of true correspondence for inference.
    Sec 3.3 defines the likelihood using CP and uses CLP in experiments; if surfaces are far apart, CLP is a poor approximation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Posterior shape models revisited: Improving 3D reconstructions from partial data using target specific models." pith.science (2026). https://pith.science/paper/LNRPNCNS

@misc{pith2026250905776,
  author       = {Pith},
  title        = {Pith review of: Posterior shape models revisited: Improving 3D reconstructions from partial data using target specific models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LNRPNCNS}},
  note         = {Machine review of arXiv:2509.05776}
}
read the original abstract

In medical imaging, point distribution models are often used to reconstruct and complete partial shapes using a statistical model of the full shape. A commonly overlooked, but crucial factor in this reconstruction process, is the pose of the training data relative to the partial target shape. A difference in pose alignment of the training and target shape leads to biased solutions, particularly when observing small parts of a shape. In this paper, we demonstrate the importance of pose alignment for partial shape reconstructions and propose an efficient method to adjust an existing model to a specific target. Our method preserves the computational efficiency of linear models while significantly improving reconstruction accuracy and predicted variance. It exactly recovers the intended aligned model for translations, and provides a good approximation for small rotations, all without access to the original training data. Hence, existing shape models in reconstruction pipelines can be adapted by a simple preprocessing step, making our approach widely applicable in plug-and-play scenarios.

Figures

Figures reproduced from arXiv: 2509.05776 by the authors.

Figure 1
Figure 1. We showcase the importance of the pose alignment of training data during partial target [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The x-axis shows the number of random targets and the y-axis shows the symmetric KL divergence to the prior for rank 3 and 45. Target-agnostic models cannot recover the prior accurately due to biased reconstructions. Formally, we are interested in the distribution of shape parameters α, rigid transformation parameters r = [rψ rγ rζ ] and t = [tx ty tz] that explain Γτ well. As such, the formulated posterior depends … view at source ↗
Figure 3
Figure 3. The right illustrates in white how much of the femur is observed for the reconstructions [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Results of the femur reconstruction experiment using MH and normalizing flows. Target [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: The predictive posterior is imposed over the target. The target-specific model leads to a [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Expected distance of the nICP reconstruction experiment. The [PITH_FULL_IMAGE:figures/full_fig_p018_6.png]
Figure 7
Figure 7. Figure 7: The parametric rotation model is a model over shapes determined by the angle [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: The left shows the recovered distribution of the angle [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Three meshes of the BFM is shown. The left mesh shows the right ear as alignment domain [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: The two meshes on the left show the point wise model variance of the target-agnostic and [PITH_FULL_IMAGE:figures/full_fig_p022_10.png]
Figure 11
Figure 11. Figure 11: The results using the scaphoid dataset. This is a repeat of the leave-one-out reconstruction [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]
Figure 12
Figure 12. Figure 12: The results of a leave-one-out femur reconstruction experiment using posterior shape [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

69 extracted references · 67 canonical work pages

  1. [1]

    Posterior shape models

    Thomas Albrecht, Marcel L \"u thi, Thomas Gerig, and Thomas Vetter. Posterior shape models. Medical image analysis, 17 0 (8): 0 959--973, 2013

    work page 2013

  2. [2]

    Planning of skull reconstruction based on a statistical shape model combined with geometric morphometrics

    Marc Anton Fuessinger, Steffen Schwarz, Carl-Peter Cornelius, Marc Christian Metzger, Edward Ellis, Florian Probst, Wiebke Semper-Hogg, Mathieu Gass, and Stefan Schlager. Planning of skull reconstruction based on a statistical shape model combined with geometric morphometrics. International journal of computer assisted radiology and surgery, 13 0 (4): 0 5...

    work page 2018

  3. [3]

    Reconstruction of 3d surfaces from incomplete digitisations using statistical shape models for manufacturing processes

    Jos \'e M Navarro-Jim \'e nez, Jos \'e V Aguado, Gr \'e goire Bazin, Vicente Albero, and Domenico Borzacchiello. Reconstruction of 3d surfaces from incomplete digitisations using statistical shape models for manufacturing processes. Journal of Intelligent Manufacturing, pages 1--14, 2022

    work page 2022

  4. [4]

    Probabilistic surface reconstruction with unknown correspondence

    Dennis Madsen, Thomas Vetter, and Marcel L \"u thi. Probabilistic surface reconstruction with unknown correspondence. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging and Clinical Image-Based Procedures, pages 3--11. Springer, 2019

    work page 2019

  5. [5]

    Virtual reconstruction of bilateral midfacial defects by using statistical shape modeling

    Marc Anton Fuessinger, Steffen Schwarz, Joerg Neubauer, Carl-Peter Cornelius, Mathieu Gass, Philipp Poxleitner, Ruediger Zimmerer, Marc Christian Metzger, and Stefan Schlager. Virtual reconstruction of bilateral midfacial defects by using statistical shape modeling. Journal of Cranio-Maxillofacial Surgery, 47 0 (7): 0 1054--1059, 2019

    work page 2019

  6. [6]

    Reconstruction of full femora from partial bone fragments for anthropological analyses using statistical shape modeling

    Lars C Ebert, Dana Rahbani, Marcel L \"u thi, Michael J Thali, Angi M Christensen, and Barbara Fliss. Reconstruction of full femora from partial bone fragments for anthropological analyses using statistical shape modeling. Forensic Science International, 332: 0 111196, 2022

    work page 2022

  7. [7]

    Determination of sex from femora

    Gita Mall, Matthias Graw, Kristina-D Gehring, and Michael Hubig. Determination of sex from femora. Forensic science international, 113 0 (1-3): 0 315--321, 2000

    work page 2000

  8. [8]

    Morphometric sex estimation from 3d computed tomography os coxae model and its validation in skeletal remains

    Mirjana Djorojevic, Concepci \'o n Rold \'a n, Patricia Garc \' a-Parra, Inmaculada Alem \'a n, and Miguel Botella. Morphometric sex estimation from 3d computed tomography os coxae model and its validation in skeletal remains. International journal of legal medicine, 128: 0 879--888, 2014

    work page 2014

Show all 69 references
  1. [9]

    Accuracy of 3d corrective osteotomy for pediatric malunited both-bone forearm fractures

    Kasper Roth, Eline van Es, Gerald Kraan, Denise Eygendaal, Joost Colaris, and Filip Stockmans. Accuracy of 3d corrective osteotomy for pediatric malunited both-bone forearm fractures. Children, 10 0 (1): 0 21, 2022

    2022

  2. [10]

    Using statistical shape models to optimize tka implant design

    Ingrid Dupraz, Arthur Bollinger, Julien Deckx, Ronja Alissa Schierjott, Michael Utz, and Marnic Jacobs. Using statistical shape models to optimize tka implant design. Applied Sciences, 12 0 (3): 0 1020, 2022

    2022

  3. [11]

    Detecting global and local hippocampal shape changes in alzheimer's disease using statistical shape models

    Kai-kai Shen, Jurgen Fripp, Fabrice M \'e riaudeau, Ga \"e l Ch \'e telat, Olivier Salvado, Pierrick Bourgeat, Alzheimer's Disease Neuroimaging Initiative, et al. Detecting global and local hippocampal shape changes in alzheimer's disease using statistical shape models. Neuroi...

    2012

  4. [12]

    Prediction of femoral head coverage from articulated statistical shape models of patients with developmental dysplasia of the hip

    Penny R Atkins, Praful Agrawal, Joseph D Mozingo, Keisuke Uemura, Kunihiko Tokunaga, Christopher L Peters, Shireen Y Elhabian, Ross T Whitaker, and Andrew E Anderson. Prediction of femoral head coverage from articulated statistical shape models of patients with developmental d...

    2022

  5. [13]

    Discrete deformable model guided by partial active shape model for trus image segmentation

    Pingkun Yan, Sheng Xu, Baris Turkbey, and Jochen Kruecker. Discrete deformable model guided by partial active shape model for trus image segmentation. IEEE transactions on Biomedical Engineering, 57 0 (5): 0 1158--1166, 2010

    2010

  6. [14]

    Shape-aware surface reconstruction from sparse 3d point-clouds

    Florian Bernard, Luis Salamanca, Johan Thunberg, Alexander Tack, Dennis Jentsch, Hans Lamecker, Stefan Zachow, Frank Hertel, Jorge Goncalves, and Peter Gemmar. Shape-aware surface reconstruction from sparse 3d point-clouds. Medical image analysis, 38: 0 77--89, 2017

    2017

  7. [15]

    Statistical biomechanical surface registration: application to mr-trus fusion for prostate interventions

    Siavash Khallaghi, C Antonio S \'a nchez, Abtin Rasoulian, Saman Nouranian, Cesare Romagnoli, Hamidreza Abdi, Silvia D Chang, Peter C Black, Larry Goldenberg, William J Morris, et al. Statistical biomechanical surface registration: application to mr-trus fusion for prostate in...

    2015

  8. [16]

    A localized statistical motion model as a reproducing kernel for non-rigid image registration

    Christoph Jud, Alina Giger, Robin Sandk \"u hler, and Philippe C Cattin. A localized statistical motion model as a reproducing kernel for non-rigid image registration. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 261--269. Sp...

    2017

  9. [17]

    Multi-resolution multi-object statistical shape models based on the locality assumption

    Matthias Wilms, Heinz Handels, and Jan Ehrhardt. Multi-resolution multi-object statistical shape models based on the locality assumption. Medical image analysis, 38: 0 17--29, 2017

    2017

  10. [18]

    A kernelized multi-level localization method for flexible shape modeling with few training data

    Matthias Wilms, Jan Ehrhardt, and Nils D Forkert. A kernelized multi-level localization method for flexible shape modeling with few training data. In Medical Image Computing and Computer Assisted Intervention--MICCAI 2020: 23rd International Conference, Lima, Peru, October 4--...

    2020

  11. [19]

    Multi-shape-hierarchical active shape models

    Juan J Cerrolaza, Arantxa Villanueva, and Rafael Cabeza. Multi-shape-hierarchical active shape models. In Proceedings of the International Conference on Image Processing, Computer Vision, and Pattern Recognition (IPCV), page 1. Citeseer, 2011

    2011

  12. [20]

    An anatomical region-based statistical shape model of the human femur

    Ju Zhang, Duane Malcolm, Jacqui Hislop-Jambrich, C David L Thomas, and Poul MF Nielsen. An anatomical region-based statistical shape model of the human femur. Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 2 0 (3): 0 176--185, 2014

    2014

  13. [21]

    Lagrangian frame diffeomorphic image registration: Morphometric comparison of human and chimpanzee cortex

    Brian B Avants, P Thomas Schoenemann, and James C Gee. Lagrangian frame diffeomorphic image registration: Morphometric comparison of human and chimpanzee cortex. Medical image analysis, 10 0 (3): 0 397--412, 2006

    2006

  14. [22]

    Symmetric diffeomorphic modeling of longitudinal structural mri

    John Ashburner and Gerard R Ridgway. Symmetric diffeomorphic modeling of longitudinal structural mri. Frontiers in neuroscience, 6: 0 197, 2013

    2013

  15. [23]

    Optimization of surface registrations using beltrami holomorphic flow

    Lok Ming Lui, Tsz Wai Wong, Wei Zeng, Xianfeng Gu, Paul M Thompson, Tony F Chan, and Shing-Tung Yau. Optimization of surface registrations using beltrami holomorphic flow. Journal of scientific computing, 50: 0 557--585, 2012

    2012

  16. [24]

    Diffeomorphometry and geodesic positioning systems for human anatomy

    Michael I Miller, Laurent Younes, and Alain Trouv \'e . Diffeomorphometry and geodesic positioning systems for human anatomy. Technology, 2 0 (01): 0 36--43, 2014

    2014

  17. [25]

    Topology-preserving shape reconstruction and registration via neural diffeomorphic flow

    Shanlin Sun, Kun Han, Deying Kong, Hao Tang, Xiangyi Yan, and Xiaohui Xie. Topology-preserving shape reconstruction and registration via neural diffeomorphic flow. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20845--20855, 2022

    2022

  18. [26]

    Statistical shape models for 3d medical image segmentation: a review

    Tobias Heimann and Hans-Peter Meinzer. Statistical shape models for 3d medical image segmentation: a review. Medical image analysis, 13 0 (4): 0 543--563, 2009

    2009

  19. [27]

    A survey on shape-constraint deep learning for medical image segmentation

    Simon Bohlender, Ilkay Oksuz, and Anirban Mukhopadhyay. A survey on shape-constraint deep learning for medical image segmentation. IEEE Reviews in Biomedical Engineering, 2021

    2021

  20. [28]

    High-level prior-based loss functions for medical image segmentation: A survey

    Rosana El Jurdi, Caroline Petitjean, Paul Honeine, Veronika Cheplygina, and Fahed Abdallah. High-level prior-based loss functions for medical image segmentation: A survey. Computer Vision and Image Understanding, 210: 0 103248, 2021

    2021

  21. [29]

    Nested sphere statistics of skeletal models

    Stephen M Pizer, Sungkyu Jung, Dibyendusekhar Goswami, Jared Vicory, Xiaojie Zhao, Ritwik Chaudhuri, James N Damon, Stephan Huckemann, and JS Marron. Nested sphere statistics of skeletal models. Innovations for shape analysis: Models and algorithms, pages 93--115, 2013

    2013

  22. [30]

    Diffeomorphic medial modeling

    Paul A Yushkevich, Ahmed Aly, Jiancong Wang, Long Xie, Robert C Gorman, Laurent Younes, and Alison M Pouch. Diffeomorphic medial modeling. In International Conference on Information Processing in Medical Imaging, pages 208--220. Springer, 2019

    2019

  23. [31]

    Statistical models of appearance for medical image analysis and computer vision

    Tim F Cootes and Christopher J Taylor. Statistical models of appearance for medical image analysis and computer vision. In Medical Imaging 2001: Image Processing, volume 4322, pages 236--248. SPIE, 2001

    2001

  24. [32]

    i3dmm: Deep implicit 3d morphable model of human heads

    Tarun Yenamandra, Ayush Tewari, Florian Bernard, Hans-Peter Seidel, Mohamed Elgharib, Daniel Cremers, and Christian Theobalt. i3dmm: Deep implicit 3d morphable model of human heads. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 128...

    2021

  25. [33]

    Statistical shape and deformation analysis: methods, implementation and applications

    Guoyan Zheng, Shuo Li, and Gabor Szekely. Statistical shape and deformation analysis: methods, implementation and applications. Academic Press, 2017

    2017

  26. [34]

    Probabilistic principal geodesic analysis

    Miaomiao Zhang and Tom Fletcher. Probabilistic principal geodesic analysis. Advances in neural information processing systems, 26, 2013

    2013

  27. [35]

    Pointr: Diverse point cloud completion with geometry-aware transformers

    Xumin Yu, Yongming Rao, Ziyi Wang, Zuyan Liu, Jiwen Lu, and Jie Zhou. Pointr: Diverse point cloud completion with geometry-aware transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12498--12507, 2021

    2021

  28. [36]

    Implicit functions in feature space for 3d shape reconstruction and completion

    Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6970--6981, 2020

    2020

  29. [37]

    Elastic shape models for face analysis using curvilinear coordinates

    Anuj Srivastava, Chafik Samir, Shantanu H Joshi, and Mohamed Daoudi. Elastic shape models for face analysis using curvilinear coordinates. Journal of Mathematical Imaging and Vision, 33: 0 253--265, 2009

    2009

  30. [38]

    Laplace--beltrami eigenvalues and topological features of eigenfunctions for statistical shape analysis

    Martin Reuter, Franz-Erich Wolter, Martha Shenton, and Marc Niethammer. Laplace--beltrami eigenvalues and topological features of eigenfunctions for statistical shape analysis. Computer-Aided Design, 41 0 (10): 0 739--755, 2009

    2009

  31. [39]

    A comprehensive survey on 3d face recognition methods

    Menghan Li, Bin Huang, and Guohui Tian. A comprehensive survey on 3d face recognition methods. Engineering Applications of Artificial Intelligence, 110: 0 104669, 2022

    2022

  32. [40]

    3d morphable face models—past, present, and future

    Bernhard Egger, William AP Smith, Ayush Tewari, Stefanie Wuhrer, Michael Zollhoefer, Thabo Beeler, Florian Bernard, Timo Bolkart, Adam Kortylewski, Sami Romdhani, et al. 3d morphable face models—past, present, and future. ACM Transactions on Graphics (TOG), 39 0 (5): 0 1--38, 2020

    2020

  33. [41]

    Morphable face models-an open framework

    Thomas Gerig, Andreas Morel-Forster, Clemens Blumer, Bernhard Egger, Marcel Luthi, Sandro Sch \"o nborn, and Thomas Vetter. Morphable face models-an open framework. In 2018 13th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2018), pages 75--82. IEEE, 2018

    2018

  34. [42]

    Unsupervised training for 3d morphable model regression

    Kyle Genova, Forrester Cole, Aaron Maschinot, Aaron Sarna, Daniel Vlasic, and William T Freeman. Unsupervised training for 3d morphable model regression. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 8377--8386, 2018

    2018

  35. [43]

    Fake it till you make it: face analysis in the wild using synthetic data alone

    Erroll Wood, Tadas Baltru s aitis, Charlie Hewitt, Sebastian Dziadzio, Thomas J Cashman, and Jamie Shotton. Fake it till you make it: face analysis in the wild using synthetic data alone. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3681--3...

    2021

  36. [44]

    Neural head avatars from monocular rgb videos

    Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nie ner, and Justus Thies. Neural head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653--18664, 2022

    2022

  37. [45]

    Im avatar: Implicit morphable head avatars from videos

    Yufeng Zheng, Victoria Fern \'a ndez Abrevaya, Marcel C B \"u hler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545--13555, 2022

    2022

  38. [46]

    Geometric morphometrics: ten years of progress following the ‘revolution’

    Dean C Adams, F James Rohlf, and Dennis E Slice. Geometric morphometrics: ten years of progress following the ‘revolution’. Italian journal of zoology, 71 0 (1): 0 5--16, 2004

    2004

  39. [47]

    Shape analysis of the basioccipital bone in pax7-deficient mice

    Joshua Cates, Lisa Nevell, Suresh I Prajapati, Laura D Nelon, Jerry Y Chang, Matthew E Randolph, Bernard Wood, Charles Keller, and Ross T Whitaker. Shape analysis of the basioccipital bone in pax7-deficient mice. Scientific reports, 7 0 (1): 0 17955, 2017

    2017

  40. [48]

    Confidence of model based shape reconstruction from sparse data

    Nora Baka, Marleen de Bruijne, Johan HC Reiber, Wiro Niessen, and Boudewijn PF Lelieveldt. Confidence of model based shape reconstruction from sparse data. In 2010 IEEE International Symposium on Biomedical Imaging: From Nano to Macro, pages 1077--1080. IEEE, 2010

    2010

  41. [49]

    Statistical shape modeling approach to predict missing scapular bone

    Asma Salhi, Val \'e rie Burdin, Arnaud Boutillon, Sylvain Brochard, Tinashe Mutsvangwa, and Bhushan Borotikar. Statistical shape modeling approach to predict missing scapular bone. Annals of biomedical engineering, 48 0 (1): 0 367--379, 2020

    2020

  42. [50]

    Accuracy of femur reconstruction from sparse geometric data using a statistical shape model

    Ju Zhang and Thor F Besier. Accuracy of femur reconstruction from sparse geometric data using a statistical shape model. Computer methods in biomechanics and biomedical engineering, 20 0 (5): 0 566--576, 2017

    2017

  43. [51]

    Generalized procrustes analysis

    John C Gower. Generalized procrustes analysis. Psychometrika, 40 0 (1): 0 33--51, 1975

    1975

  44. [52]

    Gaussian process morphable models

    Marcel L \"u thi, Thomas Gerig, Christoph Jud, and Thomas Vetter. Gaussian process morphable models. IEEE transactions on pattern analysis and machine intelligence, 40 0 (8): 0 1860--1873, 2017

    2017

  45. [53]

    On the nystr \"o m method for approximating a gram matrix for improved kernel-based learning

    Petros Drineas, Michael W Mahoney, and Nello Cristianini. On the nystr \"o m method for approximating a gram matrix for improved kernel-based learning. journal of machine learning research, 6 0 (12), 2005

    2005

  46. [54]

    u rgen D \

    J \"u rgen D \"o lz, Thomas Gerig, Marcel L \"u thi, Helmut Harbrecht, and Thomas Vetter. Error-controlled model approximation for gaussian process morphable models. Journal of Mathematical Imaging and Vision, 61: 0 443--457, 2019

    2019

  47. [55]

    Probabilistic principal component analysis

    Michael E Tipping and Christopher M Bishop. Probabilistic principal component analysis. Journal of the Royal Statistical Society: Series B (Statistical Methodology), 61 0 (3): 0 611--622, 1999

    1999

  48. [56]

    Validating bayesian inference algorithms with simulation-based calibration

    Sean Talts, Michael Betancourt, Daniel Simpson, Aki Vehtari, and Andrew Gelman. Validating bayesian inference algorithms with simulation-based calibration. arXiv preprint arXiv:1804.06788, 2018

    2018 arXiv

  49. [57]

    Linear operators and stochastic partial differential equations in gaussian process regression

    Simo S \"a rkk \"a . Linear operators and stochastic partial differential equations in gaussian process regression. In Artificial Neural Networks and Machine Learning--ICANN 2011: 21st International Conference on Artificial Neural Networks, Espoo, Finland, June 14-17, 2011, Pr...

    2011

  50. [58]

    Method for registration of 3-d shapes

    Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, volume 1611, pages 586--606. International Society for Optics and Photonics, 1992

    1992

  51. [59]

    Iterative point matching for registration of free-form curves and surfaces

    Zhengyou Zhang. Iterative point matching for registration of free-form curves and surfaces. International journal of computer vision, 13 0 (2): 0 119--152, 1994

    1994

  52. [60]

    The space of human body shapes: reconstruction and parameterization from range scans

    Brett Allen, Brian Curless, and Zoran Popovi \'c . The space of human body shapes: reconstruction and parameterization from range scans. ACM transactions on graphics (TOG), 22 0 (3): 0 587--594, 2003

    2003

  53. [61]

    Optimal step nonrigid icp algorithms for surface registration

    Brian Amberg, Sami Romdhani, and Thomas Vetter. Optimal step nonrigid icp algorithms for surface registration. In 2007 IEEE conference on computer vision and pattern recognition, pages 1--8. IEEE, 2007

    2007

  54. [62]

    Metropolis--hastings algorithms

    Christian Robert, George Casella, Christian P Robert, and George Casella. Metropolis--hastings algorithms. Introducing Monte Carlo Methods with R, pages 167--197, 2010

    2010

  55. [63]

    Markov chain monte carlo for automated face image analysis

    Sandro Sch \"o nborn, Bernhard Egger, Andreas Morel-Forster, and Thomas Vetter. Markov chain monte carlo for automated face image analysis. International Journal of Computer Vision, 123 0 (2): 0 160--183, 2017

    2017

  56. [64]

    A closest point proposal for mcmc-based probabilistic surface registration

    Dennis Madsen, Andreas Morel-Forster, Patrick Kahr, Dana Rahbani, Thomas Vetter, and Marcel L \"u thi. A closest point proposal for mcmc-based probabilistic surface registration. In Computer Vision--ECCV 2020: 16th European Conference, Glasgow, UK, August 23--28, 2020, Proceed...

    2020

  57. [65]

    Variational inference with normalizing flows

    Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530--1538. PMLR, 2015

    2015

  58. [66]

    Normalizing flows for probabilistic modeling and inference

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference. Journal of Machine Learning Research, 22 0 (57): 0 1--64, 2021

    2021

  59. [67]

    Invertible densenets with concatenated lipswish

    Yura Perugachi-Diaz, Jakub Tomczak, and Sandjai Bhulai. Invertible densenets with concatenated lipswish. Advances in Neural Information Processing Systems, 34: 0 17246--17257, 2021

    2021

  60. [68]

    Simulated annealing

    Peter JM Van Laarhoven, Emile HL Aarts, Peter JM van Laarhoven, and Emile HL Aarts. Simulated annealing. Springer, 1987

    1987

  61. [69]

    Predicting carpal bone kinematics using an expanded digital database of wrist carpal bone anatomy and kinematics

    Bardiya Akhbari, Douglas C Moore, David H Laidlaw, Arnold-Peter C Weiss, Edward Akelman, Scott W Wolfe, and Joseph J Crisco. Predicting carpal bone kinematics using an expanded digital database of wrist carpal bone anatomy and kinematics. Journal of Orthopaedic Research , 37 0...

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.