acceptodds
Desk rejected at ICLR 2027

ENF-SSM: Amortizing Anatomical Statistical Shape Modeling with Elastic Neural Flow

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

33 views

Abstract

Riemannian statistical shape analysis repeatedly computes deformation paths to estimate templates, evolve initial fields, and transfer deformations. We introduce ENF-SSM, which amortizes these computations through a shared endpoint-conditioned neural path. Building on the Spectral Meets Spatial backbone, we use an elastic regularizer derived from established bending, scaling, and shearing decompositions, together with spatial smoothness. The learned path supports a finite-step logarithm surrogate, frozen-network endpoint inversion, recursive template estimation, and deformation-based path regression. We derive the conditional bounds that separate time-discretization error, reference-path error, inverse conditioning, and endpoint attachment. These results identify the assumptions needed for operator reuse. Existing experiments on pancreas, hippocampus, and heart meshes report competitive reconstruction and lower online runtimes than the evaluated iterative implementations. They provide evidence for the computational utility of the shared path and its potential values in anatomical shape analysis.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Shape analysis

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_boGROpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes ENF-SSM for anatomical statistical shape modeling. It builds on the SmS/FUSS architecture and trains a shape interpolant with an elastic deformation energy. The model supports template estimation, endpoint inversion, deformation transfer, and path regression. Experiments use pancreas, hippocampus, and heart shapes.

Critical strengths

The paper reports competitive reconstruction results across three datasets.

Critical weaknesses
  1. Differences between the paper and code, and poor presentation. Algorithm 2 uses a midpoint on the learned path, while the code uses a coordinate average. Appendix A.6 warns that these two midpoints are not always the same and should not be used interchangeably without an error analysis, yet the code still makes this substitution.

For template estimation, Algorithm 1 uses a decreasing weight, 1/(i+1). However, in the code, direct_initial ignores the factor argument passed by ENDE_karcher_mean and instead uses the fixed weight 1/(pose_timestep+1). This does not follow the stated update rule, and it is unclear whether this function was used to produce Table 3.

Additionally, the formulation does not clearly explain how the objectives, assumptions, and update rules are connected. None of the figures in the paper are vector graphics, and their text becomes less clear at high zoom levels.

These issues make the method difficult to evaluate and leave it unclear whether the reported results support the stated algorithms.

  1. Insufficient evidence for the claimed benefits. The endpoint-inversion test in Table 4 checks consistency within ENF, not accuracy against an independent reference. Table 6 shows improved fitting after correction, but lacks a comparison with directly optimizing both endpoints. The benefit and costs of the proposed operations relative to SmS/FUSS remain unclear. These gaps limit the support for the claimed advantages.
Questions for authors

Which code produced tables? Was Table 3 generated using the function that ignores the argument factor? Are the trend and results the same when the implementation follows the stated midpoint and update rules in the paper?

What are the gains and costs remaining when using SmS/FUSS on the same downstream tasks?

Why was SmS evaluated without CD-based test-time adaptation? Why are FlowSSM and Point2SSM++ included in Table 1 but not Table 2?

Why are so many entries in Table 4 marked with dashes (--)? Do these entries mean that the methods cannot be applied, or that the settings were not tested?

Does your correction rule give better fitting results or run faster than directly optimizing both endpoints? Appendix A.5 gives the gradients for this optimization, so why does the main text say they are too hard to compute?

Were all data from each subject kept in the same split? How many subjects were used for later-visit prediction, and were they excluded from ENF training?

3 comments
AuthorsOpenReview ↗

Response to Reviewer boGR on W1

We thank the reviewer for the careful examination of the code and the detailed comparison with the paper, and for helpful suggestions on the presentations of our work. The comments on the implementation details allowed us to identify and correct several inconsistencies during the revision stage, and to clarify the intended algorithms more precisely. We address each point below.

Response to current inconsistency of codes and statements:

(a) Midpoint issue: In the revised manuscript, we explicitly claim that we adopt the Euclidean (coordinate) midpoint in Schild’s ladder, consistent with the practice in ESA (Hartman et al., 2023)(Line 29, https://github.com/emmanuel-hartman/H2_SurfaceMatch/blob/main/H2_deformation_transfer.py), a common efficient realization of parallel transport. In the revised manuscript, we changed the updating rule of Algorithm 2 into M←(Bj+S)/2M \gets (B_j + S)/2 to keep consistent with the code.

The explanation of coordinate midpoint and geodesic midpoint in the Appendix is due to some open-source Rm-SSM toolkits that operate with sparse control points (e.g., Deformetrica) or are designed around specific metrics for efficiency (e.g., Morphomatics, Geomstats) employing the more rigorous manifold midpoint. Although we adopt the coordinate midpoint of ESA for the same elastic energy and the same efficiency purpose, we discuss the intrinsic difference between the two kinds of midpoints in the paper for compactness. In the revised paper, we revised the technical parts of Deformation transfer in Sec 4.2 and Appendix A.6 (in red) for better presentation. The reported deformation transfer results of our methods are all from coordinate midpoint.

(b) Factor issue: The results in Table 3 were generated with the decreasing weight 1/(i+1)1/\left(i+1\right) stated in the paper, consistent with the discrete geodesic calculus approach (Hartman et al., 2023), and the theoretical analysis of decreasing weight 1/(i+1)1/\left(i+1\right) is provided in (Ho et al., 2013) . When preparing the anonymous release, we accidentally introduced an incorrect version of this function (fixed weight) due to the tight timeline. This was our mistake in the code dump.

In addition, the fixed-weight version could not have produced the reported results. On the pancreas cohort (N=273), holding the weight fixed at 1/(pose_step+1) with pose_step=6 would keep a large step size throughout the iterations and would essentially fail to converge, which is inconsistent with the stable templates we obtained. The actual runs used the 1/(i+1)1/\left(i+1\right) schedule.

Response to presentation of the manuscript:

In the revised manuscript, we adopted the suggestions of Reviewer ZMGV to emphasize the connection between our work and the discrete geodesic calculus framework (Rumpf & Writh, 2015), and to position the proposed framework as a neural network-based implementation of operators in discrete geodesic calculus.

Hence, in the revised manuscript, for each Rm-SSM task, we first give a general introduction to each operator with its clinical application at Operator tasks part of Sec. 3 Problem formulations, and then show its mathematical formulation and optimization procedure with additional motivations and its forms in discrete geodesic calculus. These parts are in blue in the main text of the revised manuscript. The assumption for the bound proof of each Rm-SSM operator's property is numerically verified in the Appendix accordingly, as required by Reviewer 1 and Reviewer 2. These parts are in red in the Appendix of the revised manuscript.

We believe the presentation of the manuscript are substantially improved by the above revisions. Also, we used the vector graphics of the involved figures for better visualization in the revised manuscript.

Reference:

Hartman, Emmanuel, et al. "Elastic shape analysis of surfaces with second-order sobolev metrics: a comprehensive numerical framework." International Journal of Computer Vision 131.5 (2023): 1183-1209.

Ho, Jeffrey, et al. "Recursive Karcher expectation estimators and geometric law of large numbers." Artificial intelligence and statistics. PMLR, 2013.

Rumpf, Martin, and Benedikt Wirth. "Variational time discretization of geodesic calculus." IMA Journal of Numerical Analysis 35.3 (2015): 1011-1046.

AuthorsOpenReview ↗

Response to Reviewer boGR on W2

Response to Table 6 shows improved fitting after correction, but lacks a comparison with directly optimizing both endpoints.:

To evaluate accuracy against an independent reference, in the revised manuscript we add Appendix A.15 (also required by Reviewer ZMGV and Reviewer VyXc, in red) that compares the learned logarithm of the trained ENF with a numerical logarithm obtained by minimizing the same GESM path energy. Correspondence is taken from the ENF-SSM matching module, and the numerical path is computed with the correspondence-based ESA solver (Dirichlet energy used for consistency).On the pancreas dataset the first shape is fixed as source XX and the remaining shapes are corresponded to it. For each pair (X,Y)(X,Y) we solvemin⁡γ1,…,γK−1∑k=0K−1EGESM(γk,γk+1)\min_{\gamma_1,\dots,\gamma_{K-1}}\sum_{k=0}^{K-1} E_{\mathrm{GESM}}(\gamma_k,\gamma_{k+1})with fixed endpoints and K=7K=7 (matching the c=1/7c=1/7 setting). The numerical logarithm is the first-step displacement of the optimized path; the learned logarithm is the initial field of the frozen ENF.On 272 pairs the mean L2L^2 error between the learned and numerical logarithms is 2.685. All tested pairs satisfy the error bound of Proposition 1 and the remainder condition remainder≤Mh/2\mathrm{remainder}\le Mh/2. A step-size sweep shows that the remainder decreases monotonically with KK and stays below the threshold (e.g., at K=6K=6, remainder 24.17<Mh/2=26.1524.17 < Mh/2 = 26.15).

These results confirm that the trained ENF logarithm approximates the numerical logarithm under the identical GESM energy, providing the independent reference check requested for Table 4.

Response to The endpoint-inversion test in Table 4 checks consistency within ENF, not accuracy against an independent reference:

The SmS architecture used to build the neural interpolant in ENF-SSM fixes the source shape by design, so the current framework does not directly optimize the starting point. Other Riemannian geodesic-regression methods (DCM, FCM, LDDMM, Geomstats) can optimize both endpoints simultaneously; however, as shown in Table 6 they require substantially longer runtimes and still yield higher L2L^2 errors and worse deformation metrics than ENF-SSM.

Directly optimizing Eq. (14) in Euclidean R3V\mathbb{R}^{3V} without the neural path primitive reduces to ordinary linear SSM. Following Reviewer VyXc’s request we have added these closed-form Euclidean baselines (both with fixed and free endpoints). They finish in <1 s but produce higher L2L^2 errors and larger deformation measures than ENF-SSM. While fast, they are linear methods and fall outside the Riemannian setting that is the focus of this work.

Thus the rectification step remains a practical improvement within the amortized Rm-SSM pipeline: it improves fit over the non-rectified ENF-SSM while staying far more efficient than iterative Riemannian solvers that optimize both endpoints.

Response to The benefit and costs of the proposed operations relative to SmS/FUSS remain unclear.:

The benefit of the proposed GESM energy relative to the original SmS/FUSS losses is quantified in the new ablation (Section 6.2, Table 8). We train the identical SmS architecture with the original objectives (SmS with ARAP, FUSS with Chamfer, and FUSS without the Chamfer term) versus our GESM energy, then evaluate the resulting networks on endpoint inversion because it is the basic Rm-SSM operation that underlies subsequent transport and regression.

As shown below, the GESM-trained model (ENF-SSM) consistently outperforms the vanilla-loss variants on both self-consistency and later-visit volume error:

Methodc=1/7c=2/7c=3/7c=4/7c=5/7c=6/7c=1.0VD ↓
FUSS (CD)5.65±0.363.84±0.252.31±0.191.47±0.211.31±0.221.27±0.181.26±0.150.232±0.153
SmS (ARAP)3.25±0.260.79±0.140.23±0.090.16±0.050.13±0.030.12±0.030.11±0.020.142±0.080
FUSS w/o CD2.43±0.140.52±0.130.16±0.020.13±0.020.11±0.010.09±0.010.08±0.010.113±0.058
ENF-SSM (GESM)2.25±0.150.41±0.100.11±0.010.09±0.010.07±0.010.06±0.010.05±0.010.095±0.061

Because all variants use the same frozen-network evaluation, the online computational cost remains essentially identical. The accuracy gains therefore come from the improved path primitive produced by the GESM energy. This demonstrates that the elastic training objective yields a measurably better amortized operator for the Rm-SSM pipeline.

Once trained, online costs of the proposed Rm-SSM operators are reported throughout Tables 3–6 and are substantially lower than repeated numerical solves after the one-time training.

AuthorsOpenReview ↗

Response to Reviewer boGR on Questions

Reponse to Q1: Please refer to Response to current inconsistency of codes and statements part in the response to W1. We claim that ENF-SSM results in the table and figures are produced by the codes in the current anonymous release.

Reponse to Q2: Please refer to Response to The benefit and costs of the proposed operations relative to SmS/FUSS remain unclear part in the reponses to W2.

Reponse to Q3:

Why no test-time adaption in SmS: We evaluated SmS without the Chamfer-distance test-time adaptation because, on anatomical shapes, this adaptation (as explored in FUSS) tends to collapse intermediate frames and degrade path quality. The same observation was first noted in the FUSS paper; we therefore report the plain SmS baseline for a cleaner comparison of the training objectives.

Why are FlowSSM and Point2SSM++ included in Table 1 but not Table 2: Table 1 assesses reconstruction / correspondence quality, where FlowSSM and Point2SSM++ are relevant learning-based baselines. Table 2 evaluates geometric properties of the interpolation paths (incremental strain, curvature, etc.). FlowSSM and Point2SSM++ are designed as shape reconstruction methods, and do not produce an explicit time-parameterized path under the same discrete elastic energy, so the path-wise descriptors of Table 2 are not directly applicable; they are therefore omitted from that table.

Reponse to Q4: The dashes (--) in Table 4 indicate that the corresponding methods cannot be applied to those experimental settings, not that the settings were left untested. The column c=1/7c=1/7 corresponds to the standard initial-value problem solvers supplied by LDDMM, ESA, DCM and FCM. For the other intermediate values of c, these methods do not offer a corresponding solver, so the entries are left blank. ENF-SSM can be evaluated directly at every cc by a single forward pass. This is now clarified in the table caption.

Reponse to Q5: According to the Response to The endpoint-inversion test in Table 4 checks consistency within ENF, not accuracy against an independent reference in the response to W2, compared with the Rm-SSM geodesic regression methods that can directly optimize both endpoints (DCM, FCM, LDDMM, and Geomstats), ENF-SSM using correction rule produces better results and runs faster. Compared with Linear SMM regression optimizing both endpoints with a closed-form solution, ENF-SSM produces better results. However, our focus is Rm-SSM and thus linear SSM methods with the same purpose fall outside the setting.

Gradient issue of Q5: Gradients in Eqn (31) is gradient of Eqn (14) (path regression in ENF-SSM without rectification). Its basic components, such as J0,iJ_{0,i} and JY,iJ_{Y,i}, do not have explicit formulations and are obtained by differentiation of the frozen network in a smooth local chart when using SmS architecture as a path primitive for Rm-SSM. This is the reason why we call it hard to compute and is used to explain our intuition for better understanding.

Reponse to Q6: The 80/10/10 split is used only for the unsupervised reconstruction experiment that trains the interpolant. The later-visit prediction task (Table 4) is a separate downstream evaluation: the already-trained interpolant is frozen and used to solve an initial-value problem.

Subjects with at least three imaging visits are included (Table 11 lists the counts: 410 subjects with 3 visits, 5 with 4, 7 with 5, and 54 with 6). These subjects are not required to be excluded from interpolant training, because the training objective of neural interpolant only concerns pairwise interpolation, while the prediction task queries the frozen operator to solve an initial-value-type problem. Overlap therefore does not constitute leakage for the initial-value setting.

Reviewer_VyXcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ENF-SSM amortizes repeated computations in anatomical statistical shape modeling using one endpoint-conditioned mesh interpolant. It retains the SmS/FUSS correspondence and interpolation architecture, replacing the deformation regularizer with bending, scaling, and shearing terms derived from an existing general elastic shape metric, plus spatial smoothness. On a fixed mesh chart after correspondence, the learned path supplies a finite-step initial-field surrogate; frozen-network endpoint inversion then enables initial-value evolution and Schild-style deformation transfer. The same primitive supports recursive template estimation and cross-sectional path regression with optional empirical rectification.

The analysis separates finite-difference error, reference-path discrepancy, inverse conditioning, and endpoint attachment, with an additional conditional fixed-endpoint action certificate. Experiments on pancreas, hippocampus, and cardiac meshes report improved reconstruction, useful downstream fitting, and lower online runtimes than the evaluated iterative implementations. The central contribution is a reusable computational pipeline rather than a new correspondence architecture, continuous metric, or Neural ODE flow.

Critical strengths
  1. A coherent amortization strategy with practically meaningful online savings. A single frozen interpolant is reused across several operations that normally require repeated path computations. The reported pancreas template time decreases from 220 minutes for ESA to four minutes, and Table 5 reports deformation-transfer times of 2.1 seconds versus 40.5 and 753.6 seconds for the evaluated LDDMM and ESA implementations. These are potentially valuable gains for repeated cohort analysis, even though total training cost and implementation dependence still need clarification.

  2. Useful reconstruction evidence with unusually explicit mathematical scope. Table 1 reports the lowest mean CD and S2S on all three datasets; the improvement over FUSS without CD is approximately 12%, 5.5%, and 4.1% in CD. The authors acknowledge that the architecture and underlying metric decomposition are inherited, that the initial field is a surrogate, and that template recursion and regression rectification are heuristic. This makes the contribution more assessable and avoids treating conditional bounds as unconditional geometric guarantees.

Critical weaknesses
  1. The central reuse of approximate geometric operators is insufficiently validated beyond reconstruction and self-consistency. Proposition 1 leaves the reference-velocity discrepancy uncontrolled; Proposition 2 assumes a stable local inverse; and Proposition 3 assumes local strong convexity and common fixed endpoints. These limitations are explicitly acknowledged, but the experiments do not measure the missing quantities. In Table 4, inversion of frames produced by the same interpolant primarily tests self-consistency, not recovery of a reference geodesic initial-value solution. Table 2 likewise does not identify an accurate path: small incremental strain can result from insufficient motion or endpoint mismatch, while larger displacement is not intrinsically better. The reported descriptors should be accompanied by endpoint residuals and comparison to a converged, matched-energy reference. For transfer, stretch-tensor agreement alone does not establish parallel transport of the full field; norm/inner-product preservation and refinement behavior would be more direct checks. Since approximate operators are recursively composed, these gaps affect the main contribution rather than only terminology. The coordinate-PCA evaluation in Appendix A.11, explicitly distinguished from tangent PCA, cannot close this gap.

  2. Evaluation controls do not yet isolate the contribution or establish independent downstream generalization. The 80/10/10 split in Sec. 5.1 is not specified at the ADNI subject level, and Sec. 5.4 gives neither the number of longitudinal test subjects nor their overlap with interpolant training. Appendix A.14 trains the demographic observer on most of the cardiac cohort and explicitly leaves sample independence to an unexplained “V9” audit; overlap with transfer evaluation remains unresolved. These are ambiguities, not evidence that leakage occurred. Correspondence is another confounder: the method jointly learns alignment, while downstream methods use its matched meshes in some experiments; no systematic fixed-correspondence comparison separates alignment gains from path/operator gains. Finally, the Geomstats regression baseline is described on Euclidean R^(3V), for which Appendix A.5 itself gives closed-form least-squares solutions, yet Table 6 reports 482.5 seconds without that direct baseline. The strongest computational and statistical claims therefore need better controlled comparisons, subject-level uncertainty, and explicit data provenance.

Questions for authors
  1. Geometric fidelity and inverse stability: With correspondence and final endpoints fixed, can you report learned-path action gaps and initial-field errors relative to a converged numerical solver using the same frozen energy weights, alongside endpoint attachment errors? Can you also report inversion failure rates and sensitivity to small field perturbations, and test transfer norm/inner-product preservation as the ladder resolution changes? Empirical diagnostics would suffice; I am not asking for a global convexity theorem.

  2. Independent test data: Were all ADNI visits from a subject assigned to one split? How many subjects and later visits contribute to Table 4, and were those subjects excluded from interpolant training and tuning? For Table 5's cardiac attributes, specify the transfer test set, its overlap with observer/interpolant training, the desired-label construction, and the meaning/results of “V9.” Please provide uncertainty with subjects, rather than correlated mesh pairs, as the sampling unit where applicable.

  3. Matched controls and computational cost: Under identical correspondences and endpoints, how does ENF-SSM compare with linear interpolation and a matched iterative elastic solver? For Table 6, please include the exact Euclidean solution for Eq. (14), and a fair Euclidean least-squares fit when both endpoints can vary after rectification. Does rectification still help relative to directly differentiating the frozen network as permitted by Eq. (31)? Please separate preprocessing, correspondence, training, and online inversion costs so the amortization break-even point can be assessed.

  4. Reproducible definitions: Please state how the Softmax energy weights are learned and their final values; resolve the absent heart scaling explanation; clarify whether Table 1 evaluates a correspondence-pulled target or the realized network terminal mesh; and reconcile Eq. (37)'s factor 2/[N(N−1)] with its sum over all ordered pairs i ≠ j. If the sum is indeed ordered, this factor doubles the stated average. Which definition was implemented, and does any correction change the reported comparisons?

6 comments
AuthorsOpenReview ↗

Response to Reviewer VyXc on W1 (Part 1)

We thank the reviewer for highlighting the practical online savings of the amortized pipeline, the strong reconstruction numbers in Table 1, and the explicit scoping of inherited versus new elements. In the revision we have addressed the main concerns along three lines: providing direct matched-energy numerical checks for the theoretical assumptions, clarifying data splits, and computational comparisons of Euclidean baselines.

Response to W1 on Proposition 1 leaves ..... comparison to a converged, matched-energy reference: We thank the reviewer for the detailed comments on the assumptions underlying the operator constructions and the associated empirical support. In the revised manuscript we added direct empirical measurements of the quantities to support Proposition 1 to 3. Because this part is very related to the Weakness 3 of Reviewer 1, we use the same response as follows:

Empirical verification of Proposition 1 on finite-step logarithm error (Appendix A.15):

To assess the bound, we compare the learned logarithm of the trained ENF with a reference numerical logarithm obtained by minimizing the same GESM path energy. Correspondence comes from the ENF-SSM matching module; the numerical path is computed with the correspondence-based ESA solver (Dirichlet energy used for consistency).

Experimental setup

On the pancreas dataset the first shape is fixed as source XX and the remaining shapes are corresponded to it. For each pair (X,Y)(X,Y) we solve min⁡γ1,…,γK−1∑k=0K−1EGESM(γk,γk+1)\min_{\gamma_1,\dots,\gamma_{K-1}}\sum_{k=0}^{K-1} E_{\mathrm{GESM}}(\gamma_k,\gamma_{k+1}) with fixed endpoints and K=7K=7 (matching the c=1/7c=1/7 setting). The numerical logarithm is the first-step displacement of the optimized path; the learned logarithm is the initial field of the frozen ENF. We report their L2L^2 discrepancy.Numerical results. On 272 pairs the mean L2L^2 error between the learned and numerical logarithms is 2.685. All tested pairs satisfy the error bound of Proposition 1 and the remainder condition remainder≤Mh/2\mathrm{remainder}\le Mh/2. A step-size sweep shows that the remainder decreases monotonically with KK and stays below the threshold (e.g., at K=6K=6, remainder 24.17<Mh/2=26.1524.17 < Mh/2 = 26.15).These results confirm that the trained ENF logarithm approximates the numerical logarithm under the identical GESM energy and that the finite-step bound holds on the evaluated data.

Empirical verification of Proposition 2 on inverse stability (Appendix A.16):

To assess the inverse stability of endpoint inversion, we examine the frozen ENF under controlled perturbations of the initial velocity field.Experimental setup. Using the same aligned pancreas meshes as in the Proposition 1 study, we first compute the learned logarithm with the frozen network. A relative perturbation of magnitude ε∈{0.01,0.05,0.10}\varepsilon\in\{0.01,0.05,0.10\} is added to this field, and endpoint inversion is performed. Stability is measured by the ratio of the resulting shape change to the perturbation size (empirical Lipschitz constant).Numerical results. The estimated Lipschitz constants (denormalized) are:

ε=0.01\varepsilon=0.01: 0.850±0.0700.850\pm 0.070 (max 0.997)

ε=0.05\varepsilon=0.05: 0.788±0.0410.788\pm 0.041 (max 0.919)

ε=0.10\varepsilon=0.10: 0.848±0.0330.848\pm 0.033 (max 0.944)

In all regimes the Lipschitz constant remains moderate (≈0.79\approx 0.79–0.850.85) and the shape change scales roughly linearly with ε\varepsilon. No sample shows abrupt instability. These results support that the endpoint-inversion operator of the trained ENF is locally Lipschitz stable, consistent with the assumption of Proposition 2.

Empirical support for Proposition 3 on action-gap bound (Appendix A4): Proposition 3 bounds the path error by a positive action gap between the learned path and a reference path that minimizes the same energy. We check whether this gap is positive and scales with path discrepancy on the trained models.Experimental setup. On all 272 aligned pancreas meshes we compute the learned path from the frozen ENF and a numerical reference path obtained by minimizing the identical GESM energy with fixed endpoints. For each pair we record the path discrepancy dd and the action gap ΔA=Alearned−Anum\Delta A = A_{\mathrm{learned}} - A_{\mathrm{num}}, then examine their mean values, Pearson correlation, and a quadratic fit ΔA≈ad2\Delta A \approx ad^2. Numerical results as follows:

Path discrepancy dd: 1.02±0.241.02 \pm 0.24

Action gap ΔA\Delta A: 0.66±0.360.66 \pm 0.36 (positive on average)

Pearson correlation corr(d,ΔA)=0.846\mathrm{corr}(d,\Delta A) = 0.846

Quadratic fit: a=0.168a = 0.168, R2=0.35R^2 = 0.35

The mean action gap is positive and ΔA\Delta A increases with path discrepancy. These statistics indicate that the action-gap condition of Proposition 3 moderately holds in a statistical sense on the evaluated shapes, so the bound remains informative for the trained ENF operators.

AuthorsOpenReview ↗

Response to Reviewer VyXc on W1 (Part 2)

Response to W1 on For transfer, stretch-tensor ....... rather than only terminology. :

We added numerical experiments for this part in the revised manuscript on Appendix A.6.1 in red, and they are summarized as follows:

To check whether stretch-tensor agreement is sufficient, we directly measure norm preservation of the transported vector as the ladder resolution changes.

Experimental setup: Using the same aligned pancreas meshes as in the Proposition 1–2 studies, two tangent vectors at the source are transported to the target by Schild’s ladder along the ENF interpolant. Resolutions K∈{2,4,6,8,10,12}K\in\{2,4,6,8,10,12\} are tested. For each KK we compute relnorm(K)=∣∥vK∥−∥v∥∣∥v∥\mathrm{rel_norm}(K)=\frac{\bigl|\|v_K\|-\|v\|\bigr|}{\|v\|}.Numerical results are as follows:

KK∥vK∥/∥v∥\|v_K\|/\|v\|relnorm\mathrm{rel_norm}
21.0390.039
41.0460.046
61.0790.079
81.0960.096
101.1230.123
121.1570.157

The relative norm error increases with resolution but stays below 16% for K≤12K\le 12. At the resolutions used in the main experiments the distortion remains limited, providing a direct check on the transport operator that is recursively composed in the pipeline.

Response to W1 on The coordinate-PCA evaluation in Appendix A.11, explicitly distinguished from tangent PCA, cannot close this gap.:
The coordinate-PCA evaluation in Appendix A.11 is used solely to compute the standard SSM metrics (generality, specificity and compactness) for shape reconstruction, following the linear-PCA protocol of Munsell et al. (2008). We explicitly distinguish it from tangent PCA under the elastic metric. In the revised manuscript, we removed this defensive writting to further improve readability.

AuthorsOpenReview ↗

Response to Reviewer VyXc on W2

Response to W2 on The 80/10/10 split in Sec. 5.1 is not specified at the ADNI subject level, and Sec. 5.4 gives neither the number of longitudinal test subjects nor their overlap with interpolant training.:

The 80/10/10 split applies only to the unsupervised reconstruction experiment that trains the interpolant on pairs of shapes. The ADNI longitudinal experiment is a separate downstream evaluation: the trained interpolant is frozen and used as an operator to solve an initial-value problem for future-shape prediction. Because the training loss never sees the prediction task, and the longitudinal evaluation only queries the already-trained operator, subject overlap between the two settings does not constitute data leakage. This is consistent with the overall claim of the paper: we learn a neural interpolant once and then reuse it as a path surrogate for efficient Rm-SSM-related tasks. Also, we added the summary of the ADNI hipp longitudinal data in Table 11 of the revised manuscript as follows:

Number of patients with ≥ 3 observations in the ADNI hippocampus dataset
(For example, the first column indicates that 410 patients were observed three times.)

# of observations3456
# of patients4105754

Response to W2 on Appendix A.14 trains...... not evidence that leakage occurred.:

Appendix A.14 trains the demographic observer on meshes newly generated by parallel-transport-based deformation transfer. These transferred meshes are distinct from the original representative meshes in the cardiac cohort and do not appear in the observer’s training split. Consequently, there is no overlap between the data used to train the observer and the meshes used in the transfer evaluation. “V9” was a typographical error and has been removed. We have clarified this point in the revised appendix.

Response to W3 on Correspondence is....gains from path/operator gains.:

To examine whether the learned correspondence and the interpolant can be decoupled, we replace the jointly trained alignment with correspondences produced by Point2SSM++ and then perform initial-value evolution using the same frozen interpolant. Endpoint inversion is the basic Rm-SSM operator on which the subsequent operators (parallel transport, regression, etc.) are built. The numerical check on the pancreas dataset is shown below. Under this substitution the consistency metrics degrade relative to the end-to-end ENF-SSM pipeline. The result indicates that, within our framework, correspondence and interpolation are tightly coupled: the interpolant is optimized for the alignment it learns jointly, and an externally supplied correspondence harms the quality of the derived operators. This ablation supports treating the learned path primitive as an integrated module rather than as a pure post-alignment interpolant.

Ablation on endpoint inversion with Point2SSM++ correspondences

Methodc=1/7c=2/7c=3/7c=4/7c=5/7c=6/7c=1.0
(P+Intp)2.87±0.320.69±0.170.16±0.040.13±0.040.09±0.020.07±0.020.06±0.01
ENF-SSM2.25±0.150.41±0.100.11±0.010.09±0.010.07±0.010.06±0.010.05±0.01

This part is now in the Appendix A.17 of the revised manuscript.

Response to W3 on Finally, ......without that direct baseline.:

The Geomstats baseline reported in Table 6 uses Fletcher’s geodesic regression (via the Geomstats library) on the Euclidean space R3V\mathbb{R}^{3V}. Although the optimum of Fletcher regression coincides with the ordinary least-squares solution in the Euclidean case, the implementation follows the generic iterative procedure of geodesic regression and therefore incurs a substantially higher runtime (482.5 s). Using a more complex space available in Geomstats for the same Fletcher regression would only increase the computational cost further. Because major advantage of ENF-SSM compared with other Rm-SSM is its efficiency, we adopt the cheapest geodesic regression in Geomstats toolkit.

In the updated Table 6 we also include linear interpolation and the closed-form Euclidean least-squares solutions. These produce L2L^2 errors comparable to the Geomstats result but finish in <1 s. However, because they are linear SSM methods while the focus of this work is Riemannian statistical shape modeling, their runtimes are not used as the efficiency baseline for ENF-SSM. We therefore report the Euclidean Geomstats instantiation as the most efficient Riemannian geodesic-regression baseline available in that toolkit, consistent with the paper’s emphasis on amortizing Rm-SSM operators.

Reviewer_ZMGVOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper modifies the Spectral Meets Spatial (SmS) / FUSS joint shape matching and interpolation network as a foundation for Riemannian statistical shape modeling with two additions:

  • as opposed to using ARAP or Chamfer for training the spatial models, the authors opt for a discrete first-order elastic energy along with a Dirichlet smoothing term
  • after training, they use the frozen network instead of standard operators in Riemannian shape statistics, serving as (an approximation of) the logarithmic and exponential map.

The authors theoretically assess the error of the error of the finite-step logarithm via a taylor expansion, stability of the inversion, and error w.r.t. a reference path via strong convexity. Comprehensive experiments are conducted on pancrease, ADNI Hippocampus, and heart meshes from the UK-BioBank. The authors assess quantities such as reconstruction capability, deformation transfer, and cross-sectional fitting.

Critical strengths
  • S1. The idea of using a deformation network to amortize computation of the logarithmic and exponential map is interesting. In this case it is applied to (Riemmanian) statistical shape modeling, which appears to not have been tried; however, the novelty claims need specification as detailed below.
  • S2. The authors correctly observe that using ARAP for deformations in the SMS/FUSS line of works has deficiencies, and replace this with an elastic energy.
  • S3. The resulting combination appears effective across a number of different tasks and datasets.
Critical weaknesses
  • W1: Weak link between the contributions. The authors adopt the SMS/FUSS backbone and modify this with an elastic energy from Su et al. (2020). They then propose using the velocity network as as a path primitive for Rm-SSM type pipelines. How do these two contributions synergize? From my understanding, vanilla SMS/FUSS could be used for the same types of inversions (which require only the network's first-step output) and for the operator approximations the authors construct. The authors do not test an important ablation that highlights the synergy between these two contributions: how do vanilla (ARAP) SmS or (dirichlet/chamfer) FUSS perform as amortized Riemannian operators?

  • W2. Imprecise positioning in the literature. The operator construction appears to be closely related to the variational discrete geodesic calculus of Rumpf & Wirth (2015) and Heeren et al. (2014) (i.e., a logarithm from the first step of a time-discrete path, an exponential obtained by inverting it, and Schild’s ladder from discrete midpoints). The paper cites Rumpf & Wirth in a list of references and does not cite Heeren et al. (2014) at all, which I find misleading. Positioning ENF-SSM as an amortization of this calculus would clarify both what is new and provide theoretical grounding. Moreover, there are several learned surrogates for Riemannian operators that are likewise not discussed; for instance, networks that learn initial momentum or velocity representations for LDDMM registration (Yang et al., 2017; Wang & Zhang, 2020), a learned exponential map on a thin-shell mesh space (Sassen et al., 2023), and discrete geodesic calculus on learned latent manifolds (Hartwig et al., 2025).

    On the experimental side, the authors benchmark efficiency against ESA and LDDMM, whereas for meshes already in dense correspondence, closed-form Riemannian shape spaces (von Tycowicz et al., 2018; Ambellan et al., 2021) provide explicit exponential and logarithm maps and would thus be a more apt basis for performance comparison. The SSM evaluation also omits relevant mesh-based methods, such as S3M (Bastian et al., 2023), which is also a baseline in FUSS, and Mesh2SSM (Iyer & Elhabian, 2023), which the paper cites and which was evaluated on the same ShapeWorks pancreas cohort.

    Using a single learned interpolation network to amortize all these Riemannian operators appears to be new, but without this context it is hard to judge exactly what the method adds.

  • W3: Limited theoretical support. Related to W2, the theory leaves the central approximation question largely unresolved. Proposition 1 constructs a bound on the finite-step logarithm error under the assumption that the learned path’s initial-velocity discrepancy from a reference geodesic is bounded. Proposition 2 shows inverse stability under a Lipschitz assumption. These are valid conditional statements, but these assumptions and constants are not established for the trained models in question. In Proposition 3, the authors also assume local strong convexity, which is likewise unverified. The analysis does not establish that the learned operators approximate those of the intended geometry. Clearer positioning relative to discrete geodesic calculus would help distinguish which constructions are inherited from new approximation guarantees. Moreover, measurements against numerical geodesics under the same energy would make the bounds more informative.

  • W4: Several writing issues throughout the paper:

    • Overly defensive writing at times and heavy semicolon usage, which detracts significantly from the readability of the paper. This is unfortunate, as the paper is overall appealing. There are a number of “X, not Y” instances (L277, L893, L1133) claims followed by a semicolon (L141, L153, L894)
    • Abstract L17 “existing experiments”? Are these not experiments conducted by the authors
    • “Our contributions are” appears twice, errantly on line 46.
    • “Geomstats Fletcher regression on R3V” is not a complete sentence.

References

Ambellan, F., Zachow, S., von Tycowicz, C. (2021). Rigid motion invariant statistical shape modeling based on discrete fundamental forms. Medical Image Analysis, 102178. doi:10.1016/j.media.2021.102178

Hartwig, F., Sassen, J., Braunsmann, J., Rumpf, M., Wirth, B. (2025). Geodesic calculus on implicitly defined latent manifolds. arXiv:2510.09468

Heeren, B., Rumpf, M., Schröder, P., Wardetzky, M., Wirth, B. (2014). Exploring the geometry of the space of shells. Computer Graphics Forum 33(5), 247–256. doi:10.1111/cgf.12450

Rumpf, M., Wirth, B. (2015). Variational time discretization of geodesic calculus. IMA Journal of Numerical Analysis 35(3), 1011–1046. doi:10.1093/imanum/dru027

Sassen, J., Hildebrandt, K., Rumpf, M., Wirth, B. (2023). Parametrizing product shape manifolds by composite networks. ICLR 2023. arXiv:2302.14665

von Tycowicz, C., Ambellan, F., Mukhopadhyay, A., Zachow, S. (2018). An efficient Riemannian statistical shape model using differential coordinates. Medical Image Analysis 43, 1–9. doi:10.1016/j.media.2017.09.004

Wang, J., Zhang, M. (2020). DeepFLASH: An efficient network for learning-based medical image registration. CVPR 2020, 4444–4452. CVF

Yang, X., Kwitt, R., Styner, M., Niethammer, M. (2017). Quicksilver: Fast predictive image registration – A deep learning approach. NeuroImage 158, 378–396. doi:10.1016/j.neuroimage.2017.07.008

Bastian, L., Baumann, A., Hoppe, E., Bürgin, V., Kim, H.Y., Saleh, M., Busam, B., Navab, N. (2023). S3M: Scalable statistical shape modeling through unsupervised correspondences. MICCAI 2023, LNCS. doi:10.1007/978-3-031-43999-5_44

Iyer, K., Elhabian, S.Y. (2023). Mesh2SSM: From surface meshes to statistical shape models of anatomy. MICCAI 2023, LNCS, 615–625. arXiv:2305.07805

Questions for authors

I would be inclined to raise my score if the authors can clarify the positioning of the paper as described in the weaknesses and better qualify whether the elastic regularization improves downstream operator accuracy compared with SmS/FUSS. Moreover, I expect the authors to establish stronger links between the theory and the experiments and to improve the paper's writing.

5 comments
AuthorsOpenReview ↗

Response to Reviewer ZMGV on W1 and W2

We thank the reviewer for highlighting the value of amortizing Riemannian operators with a learned path, the conditional error analysis, and the multi-dataset experiments. Below we respond to the main points on contribution synergy, literature positioning, and theoretical support.

Response to W1: We thank the suggestions of highlighting the need to clarify the synergy between the elastic training objective and the use of the network as a path primitive for Rm-SSM. To make this synergy explicit we added an ablation (Section 6.2, Table 8 in red) and summarize as follow: we take the same SmS architecture and train it with the original losses (SmS with ARAP, FUSS with Chamfer, and FUSS without the Chamfer term) versus our GESM energy, then evaluate the resulting networks on the basic Rm-SSM operation of endpoint inversion (initial-value evolution) because endpoint inversion with initial value is the basic numerical component supporting other operators. The GESM-trained model (ENF-SSM) consistently outperforms the vanilla-loss variants on both self-consistency and later-visit volume error. This shows that the elastic energy is not merely a reconstruction regularizer; it yields a measurably better path primitive for the amortized Riemannian pipeline.

Methodc=1/7c=2/7c=3/7c=4/7c=5/7c=6/7c=1.0VD ↓
FUSS (CD)5.65±0.363.84±0.252.31±0.191.47±0.211.31±0.221.27±0.181.26±0.150.232±0.153
SmS (ARAP)3.25±0.260.79±0.140.23±0.090.16±0.050.13±0.030.12±0.030.11±0.020.142±0.080
FUSS w/o CD (area preserving energy)2.43±0.140.52±0.130.16±0.020.13±0.020.11±0.010.09±0.010.08±0.010.113±0.058
ENF-SSM (GESM)2.25±0.150.41±0.100.11±0.010.09±0.010.07±0.010.06±0.010.05±0.010.095±0.061

Response to W2: We thank the reviewer for the helpful suggestions on literature positioning. We revised the related work part and contribution summary part of Introduction section, with revised parts in blue. Also, we clarified the position relative to discrete geodesic calculus on which are from discrete geodesic calculus and which are newly-developed at the end of section 4.2.

Discrete geodesic calculus: We have revised the related-work discussion to explicitly position ENF-SSM as an amortization of variational time-discrete geodesic calculus. The finite-step logarithm (first step of the discrete path), the exponential obtained by endpoint inversion, and Schild’s ladder with discrete midpoints follow the constructions of Rumpf & Wirth (2015) and Heeren et al. (2014). These works are now cited and discussed; the novelty of ENF-SSM lies in learning a single GESM-regularized neural interpolant that supplies these operators at inference time.

Learned surrogates: We have added discussion of related learned operator surrogates, including networks that predict initial momentum or velocity for LDDMM (Yang et al., 2017; Wang & Zhang, 2020), the learned exponential map on thin-shell meshes (Sassen et al., 2023), and discrete geodesic calculus on learned latent manifolds (Hartwig et al., 2025). ENF-SSM differs in that it amortizes the full path (not only the initial velocity) under an intrinsic elastic energy on the mesh, and re-uses the same frozen network for multiple Rm-SSM operators.

Experimental baselines: For meshes already in dense correspondence we include DCM (von Tycowicz et al., 2018) and FCM (Ambellan et al., 2021). In template estimation (Table 3) both are competitive in computation time but have larger average distance than ENF-SSM; in initial-value evolution (Table 4) and cross-sectional regression (Table 6) they produce higher L2L^{2} errors and substantially longer runtimes than ENF-SSM. On parallel transport / deformation transfer (Table 5), DCM achieves reasonable descriptor agreement (dR=7.93d_{R}=7.93, Δϕ=17.19\Delta\phi=17.19) at 12.4 s, while FCM is less accurate (dR=12.35d_{R}=12.35, Δϕ=40.49\Delta\phi=40.49) despite a faster 3.6 s; ENF-SSM obtains the best descriptor match (dR=5.61d_{R}=5.61, Δϕ=14.48\Delta\phi=14.48) in 2.1 s and also produces transferred shapes whose predicted attributes better match the target demographics. All the visualization results of DCM and FCM in Rm-SSM tasks are updated in Fig. 4, Fig. 6, and Fig. 7 in the Appendix, respectively. S3M and Mesh2SSM have been added to shape reconstruction baselines in Table 1, with AUC of SSM metrics updated in Fig.6 in the Appendix.

AuthorsOpenReview ↗

Response to Reviewer ZMGV on W3

Response to W3: We thank the reviewer for the comments on the theoretical support and the relationship to discrete geodesic calculus. In the revision we added direct empirical checks of these quantities on the trained models, and summerize them as follows:

Empirical verification of Proposition 1 on finite-step logarithm error (Appendix A.15):

To assess the bound, we compare the learned logarithm of the trained ENF with a reference numerical logarithm obtained by minimizing the same GESM path energy. Correspondence comes from the ENF-SSM matching module; the numerical path is computed with the correspondence-based ESA solver (Dirichlet energy used for consistency).

Experimental setup:

On the pancreas dataset the first shape is fixed as source XX and the remaining shapes are corresponded to it. For each pair (X,Y)(X,Y) we solve min⁡γ1,…,γK−1∑k=0K−1EGESM(γk,γk+1)\min_{\gamma_1,\dots,\gamma_{K-1}}\sum_{k=0}^{K-1} E_{\mathrm{GESM}}(\gamma_k,\gamma_{k+1}) with fixed endpoints and K=7K=7 (matching the c=1/7c=1/7 setting). The numerical logarithm is the first-step displacement of the optimized path; the learned logarithm is the initial field of the frozen ENF. We report their L2L^2 discrepancy.Numerical results. On 272 pairs the mean L2L^2 error between the learned and numerical logarithms is 2.685. All tested pairs satisfy the error bound of Proposition 1 and the remainder condition remainder≤Mh/2\mathrm{remainder}\le Mh/2. A step-size sweep shows that the remainder decreases monotonically with KK and stays below the threshold (e.g., at K=6K=6, remainder 24.17<Mh/2=26.1524.17 < Mh/2 = 26.15).These results confirm that the trained ENF logarithm approximates the numerical logarithm under the identical GESM energy and that the finite-step bound holds on the evaluated data.

Empirical verification of Proposition 2 on inverse stability (Appendix A.16):

To assess the inverse stability of endpoint inversion, we examine the frozen ENF under controlled perturbations of the initial velocity field.Experimental setup. Using the same aligned pancreas meshes as in the Proposition 1 study, we first compute the learned logarithm with the frozen network. A relative perturbation of magnitude ε∈{0.01,0.05,0.10}\varepsilon\in\{0.01,0.05,0.10\} is added to this field, and endpoint inversion is performed. Stability is measured by the ratio of the resulting shape change to the perturbation size (empirical Lipschitz constant).Numerical results. The estimated Lipschitz constants (denormalized) are:

ε=0.01\varepsilon=0.01: 0.850±0.0700.850\pm 0.070 (max 0.997)

ε=0.05\varepsilon=0.05: 0.788±0.0410.788\pm 0.041 (max 0.919)

ε=0.10\varepsilon=0.10: 0.848±0.0330.848\pm 0.033 (max 0.944)

In all regimes the Lipschitz constant remains moderate (≈0.79\approx 0.79–0.850.85) and the shape change scales roughly linearly with ε\varepsilon. No sample shows abrupt instability. These results support that the endpoint-inversion operator of the trained ENF is locally Lipschitz stable, consistent with the assumption of Proposition 2.

Empirical support for Proposition 3 on action-gap bound (Appendix A4): Proposition 3 bounds the path error by a positive action gap between the learned path and a reference path that minimizes the same energy. We check whether this gap is positive and scales with path discrepancy on the trained models.Experimental setup. On all 272 aligned pancreas meshes we compute the learned path from the frozen ENF and a numerical reference path obtained by minimizing the identical GESM energy with fixed endpoints. For each pair we record the path discrepancy dd and the action gap ΔA=Alearned−Anum\Delta A = A_{\mathrm{learned}} - A_{\mathrm{num}}, then examine their mean values, Pearson correlation, and a quadratic fit ΔA≈ad2\Delta A \approx a d^2. Numerical results as follows:

Path discrepancy dd: 1.02±0.241.02 \pm 0.24

Action gap ΔA\Delta A: 0.66±0.360.66 \pm 0.36 (positive on average)

Pearson correlation corr(d,ΔA)=0.846\mathrm{corr}(d,\Delta A) = 0.846

Quadratic fit: a=0.168a = 0.168, R2=0.35R^2 = 0.35

The mean action gap is positive and ΔA\Delta A increases with path discrepancy. These statistics indicate that the action-gap condition of Proposition 3 moderately holds in a statistical sense on the evaluated shapes, so the bound remains informative for the trained ENF operators.

In summary, these measurements against numerical geodesics under the same energy make the conditional bounds informative for the trained networks and clarify which guarantees are inherited from discrete geodesic calculus versus those verified for the learned operators.

AuthorsOpenReview ↗

Response to Reviewer ZMGV on W4

We thank the reviewer for the detailed comments on presentation.We corrected the above points, and revised the manuscript to reduce defensive phrasing and the frequent “X, not Y” constructions, and broke up long sentences that relied on heavy semicolon usage. The noted instances and similar passages elsewhere have been rewritten in a more direct style.

Furthermore, to improve overall readability, we revised the contribution summary in the introduction and expanded the final paragraph of the problem formulation to give a clearer explanation of the subsequent Rm-SSM operators, supplemented with concrete medical application examples. In Section 4.2 we have also added brief motivations when constructing each operator. All of these changes appear in blue in the revised manuscript.

We hope these revisions make the technical development easier to follow while preserving the original contributions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response

We sincerely thank the reviewers for their time and constructive feedback. We are encouraged by Reviewer ZMGV’s recognition of the core idea and significance of the work, and by Reviewer VyXc’s recognition of the practical performance; we also thank Reviewer boGR for the careful code-level verification that allowed timely clarifications.

Common Concerns Across Reviewers

The following issues were raised by more than one reviewer and constitute the main overlapping concerns. They are summarized below, together with the corresponding revisions in the updated manuscript:

1. Empirical support for the theoretical assumptions / operator accuracy

  • Reviewer ZMGV, W3; Reviewer VyXc, W1 & Q1; Reviewer boGR, W2
  • Revision: New matched-energy numerical comparisons and stability checks appear in red in Appendices A.15 (Prop. 1), A.16 (Prop. 2) and A.6.1 (Schild’s ladder norm preservation).

2. Synergy / incremental benefit of the GESM energy over vanilla SmS/FUSS

  • Reviewer ZMGV, W1; Reviewer boGR, W2
  • Revision: New ablation (identical SmS architecture with ARAP / Chamfer vs. GESM) is added in red in Section 6.2 and Table 8.

3. Clearer Positioning and Presentation, and new related baselines

  • Reviewer ZMGV, W2 & W4; Reviewer boGR, W2
  • Revision: Related-work reorganization (metric vs. discrete geodesic calculus), contribution rephrasing, medical examples in the problem formulation, and operator motivations in Sec. 4.2 appear in blue. Defensive phrasing sentences have been rewritten. Quantitative comparisons and visualization of new suggested baselines (S3M, Mesh2SSM, DCM, FCM, and linear SSM regression) are included in the related Tables in maintext and Figures in appendix.

4. Data split and evaluation protocol across tasks

  • Reviewer VyXc, W2 & Q2; Reviewer boGR, Q6. These parts are directly clarified in rebuttal.

5. Code–paper consistency (midpoint and template weight)

  • Reviewer boGR, W1 & Q1
  • Revision: Algorithm 2 now explicitly uses the Euclidean midpoint; Algorithm 1’s weight is confirmed as the version used for Table 3.

These five points cover the majority of the substantive comments. All corresponding changes are marked in the revised manuscript (blue for positioning and presentation, red for new experiments and results).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.