acceptodds
Under review as a conference paper at ICLR 2027

Structural Affinity: An Algorithm for Coherent Cage-Based Shape Design

Abstract

Cage-based deformation (CBD) parametrises 3D geometry through a small set of control handles, effectively enabling a variety of tasks from animation and interactive modelling to tool-based co-design, neural deformation transfer, and aerodynamic shape optimization. Many optimization approaches relying on CBD employ regularisation strategies to reduce the distortion of the deformed mesh by encouraging smooth, regular cage deformations. These techniques offer soft guidance to an optimization algorithm towards desirable deformations, but compete with the task objective rather than changing the deformation parameterization itself. In this work we introduce Structural Affinity for CBD, an explicit regularisation method which enables smooth deformations through control cages by exploiting a two-cage structure to couple cage-handle motion directly. Structural Affinity precomputes an affinity matrix using mean-value coordinates, binding the inner control-cage to a coarser outer-cage. This matrix defines how much a cage handle displacement propagates to neighbouring handles which substantially decreases and often eliminates cage self-intersections, reduces mesh distortion, and adds linear overhead at deformation time. Across robotic control–morphology co-design, source-to-target neural deformation transfer, and aerodynamic wing shape optimization, Structural Affinity produces more regular deformations and substantially reduces reliance on hand-designed structural loss terms. We report consistent gains on three complementary diagnostics: the cotangent Laplacian, a proxy for surface smoothness; cage distortion, capturing whether control handles move coherently; and surface roughness, which measures undesirable oscillation in wing geometry.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robotics, shape, geometry, cage based deformation, deformation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_taS5OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This papers introduces a new method for 3D surface mesh deformation. The method builds upon cage-based deformation and their associated cage coordinates, which generalize barycentric coordinates to polyhedron. The key idea in the paper is to use a second cage, with encloses to standard one, to scaffold the deformation and obtain better metric on the final surface deformation. The paper uses Mean Value Coordinates as the deformation engine, and The paper is evaluated on various tasks, including neural cages, aerodynamic shape optimization and robot morphology co-design. Each use case report marginal to significant gain in various metrics.

Critical strengths
  1. The paper's core idea is quite simple and effective
  2. Its computational cost seems reasonable
  3. The framework could extend to other type of cage coordinates
  4. Multiple use cases are explored for validation
Critical weaknesses
  1. It remains unclear how performance gains are maintained for other type of cage coordinates
  2. A better assessment of the 2-cage system for spiky features and sharp structure - and more generally a qualitative analysis from a geometric point of view - could help
  3. I cannot assess from the paper whether the method is sensitive to the cage structure. in other words, given two inner cages, with with different vertex density and connectivity, for the same target surface to deform, which criteria makes one better than the other?
  4. in general, the visuals of the paper do not help evaluating qualitatively the results
Questions for authors
  1. Are there any foreseeable limitations for application to other cage coordinates, namely green coordinates ?
  2. Does the method extends to quad and triquad cages and what adaptation would be needed ?
  3. In the appendices, one limitations worries me a bit (page 15): the non preservation of common translation. This is not a show stopper, but a bit buried in the document and I'd like to understand more what this implies in practice. Particularly for (interactive) design use cases.
4 comments
AuthorsOpenReview ↗

Response to Reviewer taS5

Thank you for the thoughtful review, and for recognizing that the core idea is simple and effective, that its cost is reasonable, and that the framework could extend to other cage coordinates. Your questions on exactly that extension are answered below and in the new App. A.2.

W1 / Q1: Other cage coordinates, e.g. Green coordinates.

Structural Affinity uses coordinates in two separate places.

  • Inner cage to mesh. Structural Affinity only changes the handle displacements before the mesh is deformed. Any coordinates can be used here unchanged: MVC, harmonic, biharmonic, etc.
  • Inner cage to outer cage, used to build the affinity C. Eq. 2 needs weights that are nonnegative and sum to one. MVC on a convex outer cage, harmonic coordinates and bounded biharmonic weights all qualify.

Green coordinates remain fully usable for the mesh deformation itself. For building C, their face-normal terms can be negative, so they would first be clipped or paired with a different overlap measure. This is our analysis rather than an experiment which is discussed in Sec. 7 and App. A.2.

Q2: Quad and tri-quad cages.

Eq. 2 depends only on the weight vectors, not on the face type. A quad or mixed outer cage needs only polygonal mean-value coordinates; the affinity and the propagation are unchanged (App. A.2).

W3: Sensitivity to cage structure.

  • Connectivity. C depends only on the inner-handle positions and the outer cage. Inner-cage connectivity affects only the handle-to-mesh weights, not the coupling.
  • Density. The aerodynamic study varies this on the same wing, and the benefit mostly appears at the medium and large scales, where independent handles have the most freedom to oscillate (Fig. 5).
  • Outer cage. The results are robust to a coarser and to a perturbed outer cage (App. B.4), and to the outer-cage discretization in aerodynamics (Sec. 5).

The denser of two inner cages benefits more from coupling, which intuitively follows from the smoothness-expressiveness tradeoff. We do not yet have a formal criterion for ranking two inner cages.

W2 / W4: Qualitative, geometric assessment of spiky features.

We agree the original visuals were too thin. The new Fig. 3 shows, for each test pair, the deformed mesh and the final cage of both methods. The meshes reach the target with similar fidelity; the Vanilla NC cages fold over, stretch into spikes and have irregular triangles, while the CBD-SA cages stay compact and evenly triangulated. App. B.7 adds 22 further visual comparisons (Figs. 13–16).

To isolate spikes, we also ran controlled one-handle edits on a 3D fingertip mesh (App. C, Fig. 17). Over 210 edits, CBD inverts mesh triangles in 107 and CBD-SA in 44; in the edit shown, CBD inverts 149 of 5,120 triangles and CBD-SA none. Deep inward presses can still invert triangles under CBD-SA, though far less often than under CBD.

Q3: Non-preservation of common translation.

This deserves more visibility. App. A.1 now points to a dedicated explanation in App. A.2 and A.4, and exact handle placement is stated in the method section.

  • In deployment: if every handle moves by the same vector, the coupling amplifies it. In our three applications this never arises, because global pose is fixed or optimized separately from the cage.
  • In interactive design: translation can be kept separate by propagating only the mean-free part of the displacement, or by handling pose outside the cage. A designer can still place a set of handles exactly: the propagation is linear and invertible, so the required inputs follow from a small linear solve, D = A⁻¹T.

Interactive editing is outside the scope of this paper, which targets optimizer-driven design, but we now describe how Structural Affinity can be applied there (App. A.4).

Thank you again for your thorough evaluation. It has been a pleasure to engage with your questions and comments and we hope that our new analysis addresses them. We are excited to continue discussing our work with you.

Dear authors, Thank you for you revision. It answers my questions.
I understand you do not have a formal criterion for ranking two inner cages, but in practice then, which "principles" did you apply to create them in the reported experiments?

AuthorsOpenReview ↗

Response to Reviewer taS5

Thank you very much for the follow-up. We are glad the revision answers your questions and enthused by your engagement. In every experiment we followed one principle: the inner cage is never designed for Structural Affinity. We reuse the inner cage of the baseline or of prior work, so CBD and CBD-SA always share the identical inner cage and the comparison isolates the coupling.

  • Neural Cages: the inner cage is predicted for each shape by the network, from the published 42-vertex template of Neural Cages, unchanged.
  • Aerodynamics: a closed cage that follows the wing's chordwise and spanwise stations, padded slightly around the wing, at the same lattice resolution as the FFD baseline (2×3, 4×6 and 8×12 stations for Small, Medium and Large), so that design-variable counts match FFD.
  • Robot co-design: a coarse structured cage around the fingertip (42 vertices).
  • Articulated characters (App. A.5): the human cage released with Neural Cages.

In practice, these follow the usual cage-design guidelines: enclose the shape with a small margin, follow its main sections, and give the cage as many handles as the task needs degrees of freedom. Structural Affinity then makes denser cages safer to use, since nearby handles move together rather than independently, which is where its benefit was largest.

Thank you again for your engagement, and we would be very happy to answer any further questions.

Thank you for clarifying.

Reviewer_SzygOpenReview ↗
Rating
3(+2)
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes Structural Affinity (SA), an explicit regularization method for Cage-Based Deformation (CBD) designed to produce coherent 3D shape deformations by coupling the movements of control handles. The authors identify that conventional CBD allows control handles to move independently, often resulting in distorted, spiky, or degenerate shapes. Although existing approaches address these issues through additional regularization losses, such losses can conflict with the primary optimization objective and require task-specific hyperparameter tuning. To address these limitations, the authors introduce a two-cage structure consisting of an inner control cage and a coarser enclosing outer cage. They compute Mean-Value Coordinates (MVC) between the inner and outer cages and derive a Structural Affinity Matrix by measuring the overlap between the MVC weight distributions of different inner-cage handles. This matrix determines how strongly the displacement of one handle propagates to other structurally related handles, encouraging coherent deformation without requiring additional optimization-based smoothing. The affinity matrix is precomputed and sparsified, allowing deformation propagation through a simple matrix multiplication with low computational overhead. Unlike directly constraining deformation to the outer-cage subspace, Structural Affinity preserves the full deformation space under the dense-matrix formulation with appropriate affinity strength. The authors evaluate the proposed method across neural deformation transfer, aerodynamic wing shape optimization, and robotic morphology–control co-design. In neural deformation transfer, Structural Affinity substantially reduces cage distortion and self-intersections.

Critical strengths

1. Deformation regularity through parameterization rather than additional losses.

The paper addresses a limitation of Cage-Based Deformation (CBD): independently moving control handles can produce irregular, distorted, and self-intersecting geometries. Rather than introducing additional regularization losses, the authors propose modifying the deformation parameterization itself to encourage structurally coherent handle movements.

2. Improvements in deformation regularity with limited reconstruction degradation.

Experiments on cage deformation show reductions in cage Laplacian distortion (81–84%) and mesh Cot-Laplacian distortion (26–69%) compared with the original Neural Cage. The proposed method retains most of the cage-level regularity improvements even after removing explicit shape-regularity losses, while DCD and SAUCD remain within 1% of the baseline.

Critical weaknesses

1. Insufficient evaluation of the trade-off between deformation regularity and expressiveness.

The proposed method encourages structurally related cage handles to move coherently, but such coupling may suppress localized or articulated deformations that require independent handle movements. Prior work, OptCtrlPoints [1] highlights the importance of strategically positioned control points for accurately reproducing complex shape variations, particularly around articulated joints. However, the evaluation of Structural Affinity is limited to relatively constrained shape categories in its neural deformation experiments, without examining whether the proposed coupling negatively affects large or highly localized deformations. In addition, the invertibility of the dense affinity operator does not necessarily guarantee that arbitrary deformations remain equally accessible during optimization.

2. Lack of preservation of prescribed control-handle movements.

Unlike conventional handle-based deformation methods that enforce positional constraints on control points, Structural Affinity modifies each handle's prescribed displacement by accumulating contributions from other handles. The formulation does not preserve uniform translations or satisfy multiple simultaneously prescribed handle displacements exactly, as acknowledged in Appendix A.1. Although these limitations may be acceptable in the evaluated optimization tasks, they raise concerns about the applicability of the method to general-purpose cage-based deformation and interactive shape editing.

3. Unclear advantages of the proposed structural affinity over existing geometry-aware deformation parameterizations.

The paper primarily motivates Structural Affinity as an alternative to loss-based regularization. However, established approaches based on biharmonic deformation already incorporate geometric smoothness into the deformation operator itself. OptCtrlPoints [1] demonstrates that optimizing the placement of sparse control points can improve deformation fidelity without relying on additional shape-regularity losses. Although OptCtrlPoints addresses a different aspect of handle-based deformation, the paper does not sufficiently discuss how Structural Affinity relates to such parameterization-based alternatives [1, 2, 3]. A more comprehensive comparison or discussion is essential to help clarify the methodological contribution beyond existing geometry-aware deformation techniques.

4. Limited evidence that outer-cage MVC affinity captures meaningful structural relationships.

The proposed affinity matrix is derived from the overlap of MVC weight distributions relative to an enclosing outer cage. However, the Neural Cages experiments use a fixed, generic icosahedral outer cage, which does not explicitly represent semantic part boundaries or articulation structures. Furthermore, the authors report that Structural Affinity and Euclidean Gaussian propagation share 87.5–94.2% of their selected handle relationships under the canonical configuration. Although Structural Affinity achieves better regularity in the evaluated settings, it remains unclear to what extent these improvements arise from meaningful structural information rather than a geometry-dependent smoothing prior [2, 3, 4, 5]. Experiments involving articulated objects or independently moving nearby parts would help establish the benefits and limitations of the proposed affinity formulation.

5. Insufficient qualitative evaluation of cage structures and deformation results.

Although the primary contribution of this paper is to improve the geometric coherence of cage-based deformation, the qualitative evaluation is surprisingly limited. Figure 3 provides only a single Table example, making it difficult to assess how Structural Affinity affects cage structures and the resulting mesh deformations across different geometries and deformation scenarios. The paper lacks systematic visual comparisons demonstrating how the proposed method handles large deformations, localized edits, articulated structures, and challenging cases where neighboring handles must move independently. While additional visualizations are provided for aerodynamic optimization and robotic co-design, they do not sufficiently compensate for the lack of direct qualitative analysis of cage-based deformation. Given that geometric regularity is the central motivation of the paper, more extensive side-by-side comparisons of deformed cages and meshes, including failure cases, are necessary to substantiate the claimed improvements beyond numerical smoothness metrics.

6. Weak motivation for the selected application domains and insufficient related-work coverage.

Although the paper primarily addresses a fundamental problem in cage-based shape deformation, a substantial portion of its experimental evaluation focuses on aerodynamic shape optimization and robotic morphology–control co-design rather than conventional computer graphics deformation tasks. While the Neural Cages experiments provide some graphics-oriented evaluation, they are restricted to three ShapeNet categories, and the paper does not sufficiently investigate standard deformation scenarios such as articulated shape deformation, interactive handle-based editing, or complex localized deformations. The motivation for prioritizing aerodynamic optimization and robotic co-design over these more directly relevant tasks is not adequately established. Furthermore, despite the substantial emphasis placed on these two application domains, their related-work discussions are relatively brief and do not provide a sufficiently comprehensive comparison with established domain-specific shape parameterization and regularization approaches. This makes the experimental design appear somewhat disconnected from the paper's central methodological contribution. In addition, improvements in these downstream tasks depend on multiple factors beyond geometric deformation, including optimization strategies, physical simulation assumptions, and learned control policies, making it difficult to isolate the contribution of Structural Affinity itself. A more clearly motivated experimental design, supported by stronger domain-specific related work and a broader evaluation of conventional graphics deformation tasks must be included if the paper maintains current stance.

7. Poor figure presentation and readability.

Several figures have presentation issues that make the results difficult to interpret. In Figure 3, the intended comparison is unclear, and the figure does not effectively illustrate the differences in cage structures or deformation quality. The error bars in Figure 5 are too small to distinguish, making it difficult to assess the reported variations. Figure 7 also contains overlapping text labels that reduce readability. These issues indicate insufficient attention to figure preparation and weaken the clarity and quality of the paper.

[1] Kim et al. OptCtrlPoints: Finding the Optimal Control Points for Biharmonic 3D Shape Deformation. Computer Graphics Forum, 42(7), e14963, 2023.

[2] Wang et al., Linear Subspace Design for Real-Time Shape Deformation, ACM Transactions on Graphics, 34(4), Article 57, 2015.

[3] Jacobson et al., Bounded Biharmonic Weights for Real-Time Deformation, ACM Transactions on Graphics, 30(4), Article 78, 2011

[4] Lipman et al., Green Coordinates, ACM Transactions on Graphics, 27(3), Article 78, 2008.

[5] Joshi et al., Harmonic coordinates for character articulation, ACM Transactions on Graphics, 26(3), Article 71, 2007.

Questions for authors

The paper has several substantial shortcomings that prevent it from meeting the acceptance bar of a top-tier AI conference, ICLR. Given the number and significance of the issues outlined in the Weaknesses section, I believe that substantial revisions are necessary. Therefore, I recommend that the authors address these concerns and consider submitting an improved version of the paper to another venue.

6 comments
AuthorsOpenReview ↗

Response to Reviewer Szyg (1/2)

Thank you for your detailed review. We appreciate that you identified the core contribution precisely, "deformation regularity through parameterization rather than additional losses", and noted that the regularity gains come with DCD and SAUCD within 1% of the baseline. Your review also pointed us to a body of work that we should have engaged with, and doing so has improved the paper. We have run new experiments in accordance with your concerns and we have appropriately revised the text.

W3: Relation to OptCtrlPoints, Linear Subspace Design and Bounded Biharmonic Weights.

We start here because this was the most valuable pointer in the reviews. These methods show convincingly that geometric smoothness belongs in the deformation operator rather than in a loss, which is the same conviction behind our paper, and we now cite and discuss all of them (Sec. 2, App. A.3).

  • BBW, harmonic coordinates and Linear Subspace Design shape the handle-to-mesh weights, so that each handle moves the surface smoothly.
  • OptCtrlPoints makes the elegant observation that where sparse control points sit largely determines how faithfully a shape space can be reproduced, and optimizes that placement.
  • In comparison, Structural Affinity acts on a third object that these methods leave free: the handle-to-handle coupling inside the optimizer's parameterization, which determines which combinations of handle motions are cheap to reach.

We find it exciting that the three are complementary rather than competing. Structural Affinity never touches the handle-to-mesh weights, so it can sit directly on top of biharmonic or harmonic weights. Combined with OptCtrlPoints, one would first choose where the handles go and then decide how they move together; optimizing placement and coupling jointly is a natural and exciting next step. We are grateful for the references.

Following your pointer, we also tried Structural Affinity in OptCtrlPoints' setting (SMPL template-to-target fitting from sparse landmarks). Beyond optimizer-driven shape design, articulated characters are a natural next setting for our method, and a small change to the outer cage makes it applicable to them. With the paper's generic outer cage, CBD-SA improves on CBD but not on a tuned Laplacian penalty, because a single coarse cage couples limbs that move independently. Using outer cages per-part fixes this, without changing Structural Affinity: we add one capsule per body part to a body-shaped outer cage, so the affinity becomes the average of a whole-body and a per-part affinity. With this part-aware outer cage, CBD-SA reaches lower fitting error than CBD with a tuned Laplacian loss at 16, 24 and 32 landmarks (App. A.5). Designing or learning outer cages for articulated shapes, for example from example poses as OptCtrlPoints does for control points, is a promising direction that we now name as future work.

W2 / W6: Scope, prescribed handle movements and application domains.

The paper targets optimizer-driven computational shape design. This is why aerodynamic optimization and robot co-design are central. They are the settings where independent handles are exploited to produce spiky, non-physical shapes (Figs. 8–9), and where a regularization loss must be traded off against the task objective. In each, only the deformation parameterization differs between CBD and CBD-SA: the optimizer, simulator, constraints, budget and seeds are identical, and the aerodynamic comparisons start from identical perturbed shapes. Sec. 1 now states this scope, and states that articulated animation and interactive editing are outside it. App. A.4 however, describes the three considerations under which Structural Affinity can still be applied for this.

Coupling handle motions is the mechanism behind Structural Affinity, so a prescribed displacement is deliberately redistributed. When exact positions T are required for a set of handles, the propagation is linear and invertible, so the inputs that realize them follow from a small linear solve, D = A⁻¹T (Sec. 3, App. A.2), or by disabling propagation for that subset of handles. A common translation is amplified by the unnormalized update; in all our applications global pose is fixed or optimized separately, which removes it. We agree with you that these properties are acceptable for optimization tasks and matter for interactive editing, and the revised paper now says so.

AuthorsOpenReview ↗

Response to Reviewer Szyg (2/2)

W1 / W4: Regularity against expressiveness, and whether the coupling is structural or a smoothing prior.

  • Reconstruction is preserved. DCD and SAUCD stay within 1% of Vanilla NC across categories and loss settings (Table 1), and within 1.3% even on the most nonconvex third of shapes (App. B.6).
  • Accessibility during optimization. You are right that invertibility alone does not show that deformations remain equally reachable. Our evidence is empirical: with the same optimizer, budget and constraints, CBD-SA reaches lower drag than CBD (Fig. 6), so on these problems the coupling made good deformations easier to reach, not harder.
  • Localized edits. In 210 controlled one-handle edits on a 3D fingertip (App. C, Fig. 17), CBD inverts mesh triangles in 107 edits and CBD-SA in 44.

Structural Affinity is a smoothing prior whose geometry is set by the outer cage. The question is whether the outer cage contributes more than Euclidean proximity, and we followed your suggestion of testing independently moving nearby parts (App. C, Fig. 18). On a two-prong shape, we drag one prong and measure how far the other moves (lower is better):

Outer structureOther prong movesReading
CBD (no coupling)17%Baseline
CBD-SA, one outer cage around both prongs41%Prongs coupled
CBD-SA, one outer cage per prong17%Coupling removed
Gaussian propagation (distance-based)60%Cannot separate parts

This shows both the limitation and the answer. A generic single outer cage does couple the prongs, as you anticipated. Giving each part its own outer cage makes the affinity block-diagonal and removes the coupling entirely. A distance kernel cannot express this: Gaussian propagation couples the prongs by 47–66% at bandwidths from 0.15 of the cage diagonal upward. The outer cage is thus the place where part structure enters.

On ShapeNet, Structural Affinity is more regular than Gaussian propagation at matched reconstruction on Chair and Car; on Table the two are comparable, and a part-aware outer cage is the natural way to add structure there (App. A.5).

W6: Domain-specific baselines and graphics evaluation.

Aerodynamics already includes the standard parameterization. We compare against free-form deformation (FFD), the established parameterization in aerodynamic shape optimization, at both its baseline and a matched number of design variables (Fig. 5).

New: an explicit regularization baseline (App. D.5). To separate "changing the parameterization" from "adding a smoothness loss", we added CBD with a cage-Laplacian loss of weight λ, under the Fig. 6 protocol. λ was selected on 10 exploratory seeds and then frozen for 10 held-out seeds. Mean final drag CDC_D over the held-out runs:

5% perturbation10% perturbation
CBD0.0276110.027480
CBD + Laplacian loss (λ = 10⁻³)0.0276130.027482
CBD-SA0.0272490.027342
  • CBD-SA reaches lower drag than the tuned loss in 9 of 10 held-out pairs at both magnitudes, with lower mean roughness.
  • No λ on the grid reaches CBD-SA, even when chosen in hindsight on the held-out seeds (best: 0.027539 and 0.027480).
  • Stronger regularization (λ ≥ 1) raises drag and worsens roughness.
  • The loss does what it is asked: it lowers the cage's Laplacian energy about 2.6×. The optimized wing does not improve. Penalizing the cage competes with the objective, whereas Structural Affinity changes which deformations are easy to reach.

Robotics. We did not run a loss-based baseline here: each arm costs about eight hours for each of four seeds. We consider the aerodynamic baseline above the cleaner test of the same question.

Graphics evaluation. The Neural Cages experiments follow the published protocol and its three ShapeNet categories exactly, so that our numbers are directly comparable with that work.

W5 / W7: Qualitative evaluation and figures.

We agree that one Table example was not enough for a paper whose motivation is geometric regularity.

  • Fig. 3 now shows the deformed mesh and the final cage of both methods for each test pair. The meshes reach the target with similar fidelity; the Vanilla NC cages fold over, stretch into spikes and have irregular triangles, while the CBD-SA cages stay compact and evenly triangulated.
  • App. B.7 (Figs. 13–16) adds 22 further test pairs with source, target and both results, and App. C adds the controlled cases above.
  • Fig. 5: the two "samples" behind each bar were two deterministic optimizers, so error bars were the wrong encoding. Each optimizer is now its own marker (circles SLSQP, squares IPOPT) with the mean as a bar, and the conclusions are the same under both optimizers.
  • Fig. 7: redrawn with no overlapping labels.

We hope the new experiments, the clarified scope and the discussion of the complementary methods you pointed us to address your concerns, and we would be glad to answer any further questions.

Thank you for the authors' response. I have carefully read it, along with the other reviewers' comments. Many of my previous concerns have been substantially addressed. However, I still have some remaining concerns.

I noticed that the following statement was added in Section 4, Line 230: "Figure 3 illustrates the cage improvement: the CBD-SA cages stay compact and evenly triangulated where Vanilla NC folds over or stretches into spikes, while both methods reach the target with similar fidelity."

However, upon examining the updated Figure 3, I observed that the deformed Vanilla NC mesh preserves the symmetry of the chair, whereas the CBD-SA mesh exhibits inconsistent thickness between the two armrests. I believe that the quality of the resulting deformed mesh is more important than the presence of fold-overs, spikes, or irregular triangles in the control cage itself. From this perspective, I find it difficult to agree with the authors' claim that the proposed method achieves improved deformation quality while maintaining similar fidelity.

Furthermore, even with the additional examples in Figure 13, it remains unclear where and to what extent the cages exhibit significant fold-overs, spikes, or irregular triangles. I would appreciate it if the authors could provide further clarification on these observations.

Reviewer_5RkSOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces Structural Affinity, which uses outer-cage MVC coordinates to couple inner-cage handle motion. Experiments cover neural deformation, wing optimization, and robot co-design.

Critical strengths

The method is simple, integrates readily into existing CBD pipelines, and shows substantial geometric regularity improvements. Comparisons with Gaussian propagation and direct projection provide useful ablations.

Critical weaknesses
  1. Although input objects need not be convex, the convex outer-cage prior may inadequately capture deep concavities, thin components, and independent branches. This could introduce inappropriate motion coupling. Generalization to more complex nonconvex structures remains insufficiently demonstrated.
  2. Training times and propagation complexity do not establish end-to-end inference efficiency. The costs of network inference, any required MVC/affinity recomputation, and mesh deformation are not separately reported, particularly when cages vary across inputs.
Questions for authors
  1. Could the authors evaluate more challenging nonconvex objects against CBD and Gaussian propagation, reporting deformation accuracy, distortion, and representative failures?
  2. Could the authors provide end-to-end inference timings and component breakdowns, clearly separating precomputation from per-input costs?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 5RkS

Thank you for the positive assessment. We are glad the method came across as simple and easy to integrate, and that the Gaussian and direct-projection ablations were useful. Both of your questions led to new experiments.

W1 / Q1: Nonconvex objects, thin parts and independent branches.

The existing categories are strongly nonconvex. We measure nonconvexity as 1 − (solid volume / convex-hull volume). The medians are 0.30 for Car, 0.83 for Chair and 0.86 for Table: chairs and tables are dominated by thin legs and separated supports.

The gains hold on the most nonconvex shapes. We split the test instances into three equal groups by nonconvexity and compare within each (App. B.6, Table 9). At a glance: CBD-SA cuts cage Laplacian distortion by 74–86% in every group, including shapes with mean nonconvexity 0.93.

CategoryLeast nonconvex group (mean nonconvexity)Cage Laplacian reduction (NoReg / NC Loss)Most nonconvex group (mean nonconvexity)Cage Laplacian reduction (NoReg / NC Loss)
Chair0.5680.6% / 83.1%0.9382.5% / 83.5%
Table0.6280.1% / 83.0%0.9374.4% / 79.2%
Car0.2477.4% / 83.8%0.5079.4% / 85.7%

Reconstruction accuracy (DCD) stays within 1.3% of Vanilla NC in every group.

Against Gaussian propagation. As you requested, Table 9 now also reports Gaussian propagation in every group. In the most nonconvex group it reduces cage Laplacian distortion by 78.9% on Chair, 79.4% on Table and 70.3% on Car, against 82.5%, 74.4% and 79.4% for CBD-SA NoReg. CBD-SA is more regular on Chair and Car in every group; on Table, Gaussian is slightly ahead. This matches our statement in Sec. 4 that Gaussian propagation remains a strong alternative when the outer cage adds little structure, and a more structured outer cage, such as the part-aware cage of App. A.5, is the natural way to add it.

Considerations. We report two, each with its fix. First, mesh-level smoothness is less uniform than cage regularity: without the regularity losses, the mean Cot-Laplacian does not improve in four of the nine groups (Table 9), because a few heavy-tailed pairs dominate the mean. With the original losses restored (NC Loss), it improves in eight of nine groups. Second, you are right that a single convex outer cage can couple independent branches, which is an astute observation. We show this directly on a two-prong shape: we drag one prong and measure how far the other moves, as a share of the dragged prong's motion. Lower is better, since the other prong should stay put (App. C, Fig. 18):

Outer structureOther prong movesReading
CBD (no coupling)17%Baseline leak through the cage-to-mesh weights
CBD-SA, one outer cage around both prongs41%Prongs coupled: the limitation
CBD-SA, one outer cage per prong17%Coupling removed: same as CBD
Gaussian propagation (distance-based)60%Cannot separate nearby parts

With one outer cage per part, the affinity becomes block-diagonal and the unwanted coupling disappears, while each prong keeps its internal coherence. A distance kernel cannot make this distinction. The outer cage is therefore the lever for encoding part structure: the same idea, a part-aware outer cage, lets Structural Affinity fit articulated SMPL characters better than a tuned Laplacian loss (App. A.4, A.5).

W2 / Q2: End-to-end inference cost, separating precomputation from per-input cost.

App. A.7 (Table 3) now reports this. At a glance: no added cost per optimizer iteration, and 0.02 ms per input in batches.

SettingCBD / Vanilla NCCBD-SAAdded by Structural Affinity
Aero: one-time construction of C–26 ms26 ms, once
Aero: cage-to-mesh deformation, per iteration0.008 ms0.008 ms0
Aero: total derivatives, per iteration9.5 ms9.5 ms0
Neural Cages: single input0.14 ms0.44 ms0.30 ms
Neural Cages: batch of 8, distinct cages0.31 ms0.95 ms0.08 ms per input
Neural Cages: batch of 32, distinct cages1.08 ms1.75 ms0.02 ms per input
  • Fixed cage (aero, robotics): C is built once, so every optimizer iteration costs the same as CBD.
  • Cage varies per input (Neural Cages): the inner-to-outer MVC weights and C are recomputed for every input. The Neural Cages rows are full forward passes (network, Structural Affinity, mesh deformation) for both methods. Vanilla NC's time is the network and deformation alone, so the difference is the whole per-input cost of Structural Affinity: 0.30 ms for one input, falling to 0.02 ms per input at batch size 32.

Thank you again for your valuable feedback. It has genuinely strengthened our research and encouraged our work. We hope that our new analysis has answered your questions, and we would be more than happy to continue discussing our work.

Thank you for your reply. Most of my concerns have been resolved, so I will keep the score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and the Area Chair

We thank all reviewers for their careful and constructive reviews. We were encouraged that the reviewers found our idea "simple and effective", that it "shows substantial geometric regularity improvements", and that it "integrates readily into existing CBD pipelines", all by building deformation regularity into the parameterization rather than relying on additional losses.

This is the heart of our contribution. The concerns raised regarded evidence and positioning rather than the method itself, so we have run new experiments in accordance with them and appropriately revised the text. Both have strengthened the paper, and the new results consistently support Structural Affinity.

What the new evidence shows

  1. Structural Affinity holds up on the hardest shapes. On the most nonconvex chairs and tables, cage Laplacian distortion drops by 74–86%, with reconstruction within 1.3% of Vanilla NC (App. B.6).
  2. Structural Affinity outperforms an explicit smoothness loss. In aerodynamics, a tuned cage-Laplacian loss leaves drag unchanged, while CBD-SA lowers drag in 9 of 10 held-out pairs at both perturbation levels (App. D.5).
  3. Structural Affinity extends beyond our domains. In OptCtrlPoints' SMPL setting, a part-aware outer cage lets CBD-SA fit articulated characters better than a tuned Laplacian loss at 16–32 landmarks (App. A.5).
  4. Structural Affinity is nearly free. With a fixed cage, an optimizer iteration costs the same as CBD; with a new cage per input, it adds 0.30 ms, or 0.02 ms per input in batches of 32 (App. A.7).
  5. Structural Affinity's benefit is visible. Fig. 3 and 22 further test pairs show compact, evenly triangulated cages where Vanilla NC folds over and spikes (App. B.7); in one-handle edits, mesh inversions fall from 107 to 44 of 210 (App. C).
  6. Structural Affinity complements existing methods. We thank Szyg for pointing us to BBW, harmonic coordinates, Linear Subspace Design and OptCtrlPoints, which broadened our view: they shape the handle-to-mesh weights or the handle placement, while Structural Affinity couples the handles themselves, so they combine naturally (Sec. 2, App. A.3).

Considerations, and how the framework handles them

  • The outer cage is where structure enters, so it can also keep parts apart. With a single convex outer cage, nearby parts move together; giving each part its own outer cage restores their independence, and a part-aware outer cage enables the articulated results above (App. A.4, A.5, C).
  • Precise control remains available. Coupling shares each displacement with related handles by design; when exact positions are needed, a small linear solve provides them (Sec. 3, App. A.2).
  • Structural Affinity works with other coordinates. Any coordinates can deform the mesh, and any nonnegative weights, such as BBW or harmonic coordinates, can build the affinity; Green coordinates would first be clipped (App. A.2).
  • Structural Affinity combines with existing losses. Where maximum surface smoothness is the goal, shape-regularity losses add to it naturally (Sec. 4).

Revisions in the PDF (in blue)

New:

  • nonconvexity analysis with Gaussian propagation (App. B.6, Table 9);
  • end-to-end inference timing (App. A.7, Table 3);
  • explicit-regularization baseline (App. D.5);
  • systematic qualitative comparisons (Fig. 3, App. B.7);
  • controlled synthetic analyses (App. C);
  • part-aware outer cages for articulated shapes (App. A.5, Table 2).

Clarified:

  • the relation to parameterization-based methods (Sec. 2, App. A.3);
  • scope and practical use, including exact handle placement, translation and other coordinates (Sec. 1, Sec. 3, Sec. 7, App. A.2, A.4).

Figures: Fig. 5 shows each optimizer separately and Fig. 7 is redrawn.

We hope we have addressed the questions and feedback raised by our reviewers, which have ultimately strengthened our scientific analysis and contribution, and shown that Structural Affinity reaches further than we originally demonstrated. For this and your time and discussion, we thank you all.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.