When we look at two distinct 3D shapes—such as two human heart ventricles or two primate teeth—our brains can instantly spot similarities and differences. We notice if one is elongated, if a curve is sharper, or if a structural feature is missing.
Historically, translating this intuition into computer algorithms required establishing explicit point-to-point correspondences. To compare shape $A$ to shape $B$, an algorithm typically needed to know exactly which vertex on mesh $A$ matched which vertex on mesh $B$. This required manual landmarking (e.g., placing digital pins on specific anatomical structures) or computationally expensive mesh registration algorithms designed to warp one surface onto another.
As highlighted by Halimi et al., this reliance on external references or rigid templates makes traditional frameworks incredibly brittle. What happens when your data contains topological noise, varying vertex counts, or holes? What if you are comparing shapes so morphologically diverse that no common template exists? Pierson et al. note that traditional template-based alignment frequently fails when confronted with the messy, unaligned realities of raw scan data.
To solve this, a powerful paradigm has emerged: correspondence-free shape analysis. Instead of treating shapes as ordered lists of discrete points or rigid geometric meshes, this approach treats shapes as continuous geometric measures or distributions. By shifting the mathematical perspective from point-matching to measure-comparison, researchers can analyze, register, and classify complex geometries without ever defining a single landmark.
The Core Intuition: Shapes as Spatial Distributions
To understand why this works, imagine a 3D surface not as a collection of connected vertices, but as a continuous physical distribution—like a cloud of dust or a localized field.
If you want to compare two clouds of dust, you don't try to pair up individual dust motes. Instead, you look at the density and layout of the clouds globally. In geometric shape analysis, this concept is formalized using mathematical constructs known as currents and varifolds.
Mathematical Currents
Instead of discrete point sets, the authors of A statistical shape modelling framework to extract 3D shape biomarkers from medical imaging data represent anatomical shapes as distributions of features using mathematical currents. A current treats a shape by how it interacts with smooth vector fields, effectively capturing both the local position and the oriented surface normal. This allows for the comparison of complex or irregular structures without requiring manual landmarking or explicit point-to-point correspondence between subjects.
Varifolds
While currents are highly effective, they can struggle when a shape's orientation is ambiguous or when dealing with unoriented geometries. This is where varifolds come in. As explained by Pierson et al. and Hartman et al., varifolds represent 3D surfaces as mathematical measures that capture both local position and local direction (tangent spaces) without relying on specific mesh connectivity or vertex order.
Traditional Mesh Alignment:
Mesh A (Vertex 1) ---> Mesh B (Vertex 1) [Requires exact matching]
Mesh A (Vertex 2) ---> Mesh B (Vertex 2) [Fails if vertex counts differ]
Varifold / Current Comparison:
Mesh A ---> Embedded as continuous measure \
+---> Inner Product in Hilbert Space
Mesh B ---> Embedded as continuous measure / [No point-to-point matching]
By embedding shapes as varifolds into a common reproducing kernel Hilbert space (RKHS), the problem of shape comparison is transformed. Evaluating the similarity between two entirely different meshes reduces to computing a kernel-based similarity measure—essentially an inner product between two geometric measures. As Pierson et al. demonstrate, this allows algorithms to evaluate geometric similarity directly on raw, unaligned data, bypassing mesh registration entirely.
This representation is incredibly versatile. For example, Durmus et al. adapted oriented varifold distances to time-series analysis. By treating a time-series signal as a geometric graph (a distribution of oriented segments), their framework achieves immense robustness against irregular sampling and missing data without needing to align time steps sequentially.
Alternative Paths to Throwing Away Landmarks
The "shape-as-measure" philosophy manifests in several other landmark-free representations, each optimizing for different types of geometric data:
- Level Sets: Instead of tracking explicit surface boundaries, Yang et al. utilize level set representations for statistical shape analysis. By embedding shapes as the zero-isocontour of a higher-dimensional function, they capture complex topological variations automatically, making it highly effective for segmenting noisy medical images.
- Global Distributions of Local Properties: Rather than aligning 3D structures in space, Fang et al. represent flexible shapes (like bending proteins) as global distributions of a "local-diameter" descriptor. This reduces the comparison of complex 3D geometries to a simple distance calculation between one-dimensional histograms, eliminating the need for spatial superposition.
- Topographic Landscapes: In biological applications where shapes lack consistent biological landmarks (such as highly varied mammalian teeth), the authors of Automated 3D Phenotype Analysis Using Data Mining treat surfaces as topographic landscapes. This allows them to extract quantitative features and classify shapes objectively without human bias.
- Integral Geometry and Topology: Bruce Wang et al. developed a pipeline called SINATRA that combines integral geometry with Gaussian process models to automatically highlight specific physical regions that differentiate classes of shapes. Taking a purely topological route, Meng et al. introduce a framework based on the Smooth Euler Characteristic Transform (SECT), providing a landmark-free and diffeomorphism-free foundation for statistical hypothesis testing on random shapes.
Statistical Modeling and Deformations: Warping Space, Not Points
If we do not have point-to-point correspondences, how do we perform statistical operations on shapes? For instance, how do we calculate an "average" shape (an atlas) from a population, or model how a shape deforms over time?
The answer lies in diffeomorphic shape analysis (or morphometry). Instead of dragging individual vertices around, these methods model shape variations by deforming the underlying space in which the shapes live.
Jia et al. utilize this approach by treating shape variations as the flux of a continuous 3D vector field. By modeling shapes as deformations of a common template, they can capture subtle anatomical differences without requiring pre-defined anatomical markers.
Point-Dragging Deformation (Traditional):
(Point A) ----> (Warped Point A)
(Point B) ----> (Warped Point B) [Can cause mesh tangling/tearing]
Space-Warping Diffeomorphism (Correspondence-Free):
[ ~ ~ ~ ~ ] The entire ambient space flows smoothly.
[ ~ Shape ~ ] --> Any geometric measure embedded inside
[ ~ ~ ~ ~ ] flows along with it, preserving topology.
This space-warping mechanism is highly scalable. Fei Wang et al. leverage information-theoretic measures of divergence to compare and simultaneously register multiple unlabeled point sets with unequal point counts, constructing an atlas automatically.
When the data becomes multi-modal, the measure-theoretic approach shines. Miller et al. extended the Large Deformation Diffeomorphic Metric Mapping (LDDMM) framework to spatial transcriptomics. By treating cells as varifolds that combine both spatial coordinates and high-dimensional genetic feature labels, they compute smooth, non-rigid transformations that align complex tissue slices across entirely different scales without requiring one-to-one cellular correspondences.
Learning the Flows
More recently, deep learning has been integrated into this framework to learn these spatial transformations directly from data. Lüdke et al. introduced FlowSSM, which uses neural networks to model continuous deformation flows. This landmark-free statistical shape modeling approach learns complex anatomical variations without requiring dense training correspondences, outputting discriminative latent representations directly from raw geometries. Similarly, Song et al. demonstrate that dictionary learning can operate directly within a shape space without pre-alignment, generating highly realistic shape reconstructions and "dictionary atoms."
Summary
Correspondence-free shape analysis works because it redefines what a shape is mathematically. By moving away from rigid, ordered vertex lists and embracing continuous geometric measures—like varifolds, currents, and spatial distributions—the field has bypassed the bottleneck of manual landmarking and template matching. Whether through kernel inner products in a Hilbert space or smooth diffeomorphic flows of the ambient environment, these methods allow us to compare, deform, and analyze the true underlying geometry of data cleanly and robustly.
Key Foundational Work
3D Image Segmentation of Deformable Objects with Shape-Appearance Joint Prior Models (2003) — contributed 1 concept elements to this direction.
Hierarchical Organization of Shapes for Efficient Retrieval (2004) — contributed 1 concept elements to this direction.