MorphModes: Non-rigid Registration via Adaptive Skinning Eigenmodes Volume: 45 Issue: 2

G. Browne* 1 and M. Liu* 1 and E. Grinspun 1 and O. Benchekroun 1 1 University of Toronto · 2025

Read MorphModes: Non-rigid Registration via Adaptive Skinning Eigenmodes Volume: 45 Issue: 2 by G. Browne* 1 and M. Liu* 1 and E. Grinspun 1 and O. Benchekroun 1 1 University of Toronto (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-18658.txt.

Introduction

Gabrielle Browne, with Mengfei Liu, Eitan Grinspun, and Otman Benchekroun, improves non-rigid registration, the task of aligning two shapes when one has bent or stretched. Registration matters in medical imaging, robotics, vision, and entertainment. Standard algorithms inch a shape toward its target and easily get stuck in poor local matches, depending heavily on where they start.

MorphModes reframes the problem as matching signed distance functions, which describe a whole shape's geometry rather than just surface points, giving the optimizer richer information. Deformations are expressed through a smooth set of skinning eigenmodes, motion patterns that prevent degenerate, crumpled solutions. An adaptive step then enlarges the pattern set where fine local bending is needed. The result is an algorithm more robust than the standard approach and less touchy about parameter settings than other distance-based methods.

Catalog record

CREATORG. Browne* 1 and M. Liu* 1 and E. Grinspun 1 and O. Benchekroun 1 1 University of Toronto
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE35 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-18658

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MorphModes: Non-rigid Registration via Adaptive Skinning Eigenmodes Volume: 45 Issue: 2 G. Browne* 1 and M. Liu* 1 and E. Grinspun 1 and O. Benchekroun 1 1 University of Toronto Abstract Non-rigid registration is a crucial task with applications in medical imaging, industrial robotics, computer vision, and entertainment. Standard approaches accomplish this task using variations on the Non-Rigid Iterative Closest Point (NRICP) algorithms, which are prone to local minima and sensitive to initial conditions. We instead formulate the non-rigid registration problem as a Signed Distance Function (SDF) matching optimization problem, which provides richer shape information compared to traditional ICP methods. To avoid degenerate solutions, we propose to use a smooth Skinning Eigenmode subspace to parameterize the optimization problem. Finally, we propose an adaptive subspace optimization scheme to allow the resolution of localized deformations within the optimization. The result is a non-rigid registration algorithm that is more robust than NRICP, without the parameter sensitivity present in other SDF-matching approaches. ccs Computing methodologies Non-Rigid Registration † † year: 2026 † † year: 2026 † † editors: B. Masia and J. Thies † † editors-short: I. Lim and P. Musialski † † editors-star: A. Bönsch and N. Umetani † † editors-tutorial: R. Chaine and A. Vaxman † † editors-education: B. Heinemann and B. Wünsche † † editors-poster: T. Gerrits and M. Teschner † † editors-dc: N.N. † † subject: EUROGRAPHICS CGF Vol No † † teaser: We introduce a Non-Rigid Registration algorithm that minimizes the SDF error between a source and a target shape. The optimization is carried out using a novel adaptive Skinning Eigenmode subspace, leading to smooth, stable morphing of one lion pose to another. *Both authors contributed equally to this research. 1 Introduction How does one shape transform into another? Registration is the process of finding an answer to this question. To an artist, this might seem like a simple task. For example, mapping between a human body and a quadrupedal creature has an obvious solution: the head of the human transforms into the head of the creature, the legs to the back legs and the arms to the front legs. However, coming up with an algorithmic solution to this problem is much more challenging. The de-facto approach to perform registration is via the Iterative Closest Point (ICP) algorithm [ BM92 ] . Unfortunately, despite its widespread use, ICP suffers from well-known limitations. Most notoriously, ICP, especially its non-rigid variant, is prone to severe local minima and null space problems that arise from the closest point projection that occurs at every one of its iterations; so long as a point on the source shape lies on the target shape, its contribution to the ICP energy will be minimized. Figure 1: Didactic illustration of the shrinkage occurring throughout non-rigid ICP iterations leading to incorrect registration. Figure 2: An example showing limb shrinkage occurring with ICP-based registration. Directly using NRICP leads to extreme shape collapse. Aligning the two meshes first using RICP, followed by NRICP leads to a better result, but with the back right leg of the hippo collapsing onto the belly. Our method leads to the correct registration. This null space problem manifests itself as "shrinkage" in the registration, where regions of the source shape are pulled and scaled down to collapse degenerately to a point or sliver on the surface of the target shape, as shown didactically in Fig. 1 . The popular fix to this null-space and non-smoothness problem is to use a regularizer that encourages smoothness and non-degeneracy in the registration. Unfortunately, this regularizer ends up combating the distance objective, sacrificing the quality of the registration for an overly smooth and non-degenerate solution as shown in Fig. 13 . To address these issues, we depart from the closest-point projection definition of the ICP objective, and instead propose using an objective that measures the difference between the volumetric signed distance function (SDF) of the source and target shapes. This objective sidesteps the crude closest point projection present within the ICP algorithm altogether, avoiding the shrinkage problem. However, this objective alone is not enough to guarantee a non-degenerate solution: it opens the door for a new null-space, where the source surface may collide, twist and fold into itself however it wants, so long as the SDFs of the source and target shapes match. Figure 3: A didactic illustration of the null space problem in the SDF matching objective, which shows that twisting and collapsing the source surface onto itself leads to the same SDF as the true solution, so long as other parts of the source shape stretch to the target shape to compensate. To avoid this new degeneracy, we propose to reparameterize our optimization problem via Skinning Eigenmodes [ BZC∗23 ] , a deformation subspace for representing smooth, energy efficient, non-rigid deformations. Introducing the subspace instead of adding a regularizer allows the optimization problem to proceed more stably without the two objectives fighting each other. In order to resolve global transformations as well as localized deformations, we specifically propose an adaptive subspace optimization scheme, that proceeds by iteratively adding more skinning eigenmodes to the subspace and resolving the registration problem at each stage. The result is a novel non-rigid registration algorithm that makes use of an adaptive skinning eigenmode subspace. Our method is more robust to initial conditions than standard ICP-based appraoches, capable of resolving registrations with large deformation, without the objective clash that arises between the competing objectives in existing regularized registration approaches. We validate our approach through comprehensive experiments across diverse scenarios, including human-to-human pose transfer, large deformation registrations between different animals in different poses, and challenging cross-category registrations, and meshes with different topologies. Our evaluations demonstrate that our method outperforms traditional non-rigid ICP approaches, particularly in scenarios requiring substantial shape deformations where ICP-based methods fail. Our experiments also highlight the robustness of our method to poor mesh quality and its ability to handle cases with no clear semantic correspondence between source and target shapes. 2 Related Work 2.1 Non-Rigid Registration The problem of registering two shapes has been studied for decades. The most classical approach to solving this problem is Iterative Closest Point (ICP), originally introduced by [ BM92 ] for medical images. Despite decades of literature on variations to this technique, the core of the method remains largely unchanged; it iteratively computes closest-point correspondences between two shapes and estimates the rigid transformation that minimizes the distances between these correspondences. Originally, only the Rigid Transformation that aligned one shape to another was sought, yielding Rigid-ICP(RICP). This method was widely applied in medical imaging [ AFP00 ] , robotics [ LM94 ] , and computer graphics [ RL01 ] with modern variants improving computational efficiency and robustness to outliers [ BTP13 , Bou15 ] . While rigid registration is well-understood, non-rigid registration is far more challenging. The solution space becomes high-dimensional, and standard ICP extensions suffer from numerous local minima, null spaces and noise sensitivity [ Sah20 ] . Early approaches simply adapted ICP with smoothness or rigidity priors or curvature-aware regularization [ ARV07 , THYS22 , LSP08 , CR03 , PCM07 ] . While these methods provide solutions to the non-rigid registration problem, ICP-based techniques still rely heavily on a good initial alignment and are

Read the complete text