SVarM: Linear Support Varifold Machines for Classification and Regression on Geometric Data

Emmanuel Hartman, Nicolas Charon · 2025

Read SVarM: Linear Support Varifold Machines for Classification and Regression on Geometric Data by Emmanuel Hartman, Nicolas Charon (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-2506-01189.txt.

Introduction

Emmanuel Hartman and Nicolas Charon build learning tools for data made of shapes in a 2025 study. Each observation might be a curve, graph, or surface, and shape spaces do not behave like flat number spaces. A model must also ignore how a shape is parameterized, or it will fail on the same shape described differently.

Their method represents shapes as varifolds, measures that pair position with direction. Duality with test functions turns shape comparison into operations a linear machine can use. Support vector style learning then handles classification and regression on geometric data. Keeping the right invariances should help the learned rules transfer to new shapes.

Catalog record

CREATOREmmanuel Hartman, Nicolas Charon
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE58 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-01189

Browse the full library · About the Tower of Babel

Full text

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

SVarM: Linear Support Varifold Machines for Classification and Regression on Geometric Data Emmanuel Hartman, Nicolas Charon Abstract Despite progress in the rapidly developing field of geometric deep learning, performing statistical analysis on geometric data—where each observation is a shape such as a curve, graph, or surface—remains challenging due to the non-Euclidean nature of shape spaces, which are defined as equivalence classes under invariance groups. Building machine learning frameworks that incorporate such invariances, notably to shape parametrization, is often crucial to ensure generalizability of the trained models to new observations. This work proposes SVarM to exploit varifold representations of shapes as measures and their duality with test functions h : ℝ n × S n − 1 → ℝ h:\mathbb{R}^{n}\times S^{n-1}\rightarrow\mathbb{R} . This method provides a general framework akin to linear support vector machines but operating instead over the infinite-dimensional space of varifolds. We develop classification and regression models on shape datasets by introducing a neural network-based representation of the trainable test function h h . This approach demonstrates strong performance and robustness across various shape graph and surface datasets, achieving results comparable to state-of-the-art methods while significantly reducing the number of trainable parameters. 1 Introduction Shape analysis, or geometric data science, is a field dedicated to building statistical and machine learning methods able to retrieve and analyze the morphological variability in geometric structures. This is a particularly central problem in applications such as computer vision or biomedical imaging where observations often come as segmented curves, surfaces, densities or other types of complex geometric data. Various approaches have been proposed, including Riemannian and elastic shape space models [ 21 , 36 , 27 , 51 ] , topology based methods [ 20 , 13 , 18 ] , and metric/functional matching frameworks [ 9 , 35 , 37 ] . With the recent advances in acquisition protocols and the explosion in the size and resolution of datasets that followed, one can expect deep learning to eventually provide tools that better scale up to modern data dimensionality. Despite this, the very nature of geometric data poses unique challenges when it comes to designing robust neural network models to tackle shape analysis tasks, which is the object of active research within the relatively new area known as geometric deep learning , see the surveys [ 10 , 12 , 19 ] and references therein for the current state-of-the-art. In this work, we introduce a shape learning framework based on a different construction than aforementioned references. The main idea is to leverage the concept of varifolds from geometric measure theory [ 2 , 1 ] along with the natural representation of a shape q\in\mathcal{S} —such as a curve or a surface in \mathbb{R}^{n} —as its corresponding varifold \mu_{q}\in\mathcal{V} . The mapping q\mapsto\mu_{q} is invariant under reparameterization, thereby providing a (nonlinear) embedding of the shape space \mathcal{S} into \mathcal{V} , the space of measures on \mathbb{R}^{n}\times S^{n-1} . This enables us to instead build the learning framework on the space of varifolds by exploiting the duality of these measures with test functions. We here simply propose to rely on linear (or affine) forms of \mathcal{V} of the form \langle\mu,h\rangle , where h\in C_{0}(\mathbb{R}^{n}\times S^{n-1}) is a trainable test function. Due to its analogy with support vector machines (but here considered in the infinite-dimensional varifold space), we coin our model S upport Var ifold M achines (SVarM) and derive formulations of the shape regression and classification problems based on this setting. Importantly, this approach allows for a principled handling of complex shapes without requiring prior regularization or registration. Additionally, by focusing on approximating continuous functions in a low-dimensional space, the model remains lightweight, requiring significantly fewer trainable parameters than traditional models operating on raw geometric data as inputs. The presentation and contributions of the paper are organized into three main sections. Section 2 introduces the mathematical settings of shape spaces and of varifolds. We then proceed by deriving a few theoretical results related to the approximation properties of SVarM for regression and classification as well as its stability with respect to geometric perturbations. Section 3 details our proposed neural network architecture, which approximates the trainable test function h of the model. Finally, Section 4 demonstrates the practical effectiveness of the approach on various regression and classification problems for real surface and shape graph datasets. Figure 1 : Representation of the quotient shape space \mathcal{S} (left) and its embedding in varifold space \mathcal{V} (right). An linear form \mu\mapsto\langle\mu,h\rangle used by SVarM is visualized as the red hyperplane in \mathcal{V} . 2 Mathematical foundations and results 2.1 Shape Space The concept of ”shape” encompasses a wide variety of geometric structures (see e.g. [ 51 , 5 , 44 ] for general overviews on shape analysis). Although the framework we introduce can in principle apply to any type of shapes that can be viewed as measures or varifolds, our focus here will be on open/closed curves embedded in the plane or 3D space, surfaces in \mathbb{R}^{3} , and by extension structures known as shape graphs [ 46 , 3 ] . The challenging aspect of dealing with such objects is that there is generally not a simple nor canonical way to represent them in a common Euclidean or even functional space. Indeed, a curve or surface is usually defined by a parametrization function q:M\rightarrow\mathbb{R}^{n} , where n=2 or 3 and M is the parameter manifold (for instance an interval or S^{1} for open and closed curves, a disk or S^{2} for open/closed surfaces \ldots ). In this paper, we will consider rectifiable shapes where q is assumed to be a bi-Lipschitz regular embedding from M to \mathbb{R}^{n} , denoted by q\in\mathcal{I}\doteq\operatorname{Emb}_{\text{Lip}}(M,\mathbb{R}^{n}) . However, q itself does not uniquely identify the shape q(M)\subseteq\mathbb{R}^{n} . The action of any reparametrization q\circ\gamma , where \gamma\in\mathcal{D}=\operatorname{Diff}_{\text{Lip}}(M) is a bi-Lipschitz homeomorphism of M , leaves the shape q(M) invariant. In other words, the actual shape associated to q is in essence the equivalence class [q]=\{q\circ\gamma:\ \gamma\in\operatorname{Diff}_{\text{Lip}}(M)\} , and the shape space \mathcal{S} can be thus identified with the quotient \mathcal{S}=\mathcal{I}/\mathcal{D} , see Figure 1 for an illustration. Note that quotienting out other shape invariances, such as to rigid motions of \mathbb{R}^{n} , can also be relevant in certain situations, but this work focuses on creating a parametrization-invariant model, as rigid alignment can usually be applied in a separate step. In fact, Section 4.1 and Section 4.2 illustrates how SVarM can itself be used for that very task. We want to add that the above shape space setting also comprises discrete objects, e.g. polygonal curves and triangulated surfaces. These correspond to the case when M is subdivided into a 1- or 2-dimensional mesh and the parametrization q is a Lipschitz piecewise affine map over this mesh, in which case q(M) can be seen as a reunion of adjacent cells in \mathbb{R}^{n} . The issue of parametrization invariance still remains very relevant in this situation. Namely, a proper learning model should exhibit some degree of robustness to operations such as mesh subdivision, simplification or resampling. 2.2 Varifolds The concept of varifold goes back to early works in geometric measure theory, notably [ 2 , 1 ] , where it provided a natural and convenient setting to

Read the complete text