Intrinsic Geometry for Non-Rigid Shape Matching

87 concepts67 papers1 proposal

Imagine a 3D model of a human standing with arms outstretched versus the same human clapping their hands. If we measure the distance between their palms using standard Euclidean coordinates, the distance drops from two meters to zero. To a traditional coordinate-based algorithm, these two poses look like completely different objects.

As noted by Alexander Bronstein et al. in A correspondence-less approach to matching of deformable shapes, non-rigid deformations alter extrinsic coordinates while preserving the object's core identity. This limitation makes standard 3D descriptors highly sensitive to physical pose or deformation state, a challenge highlighted by Ren-Hong Wang et al. in Editorial.

To recognize deformable or articulated objects reliably, computer vision must shift its perspective from extrinsic geometry (how a shape is embedded in 3D space) to intrinsic geometry (properties inherent to the shape itself). This paradigm shift, advocated by Leonidas Guibas et al. in Intrinsic Multiscale Geometry, relies on the concept of isometry. An isometric transformation is one that bends or twists a surface without stretching or tearing it. Under pure isometry, the shortest path between any two points along the surface remains exactly the same, regardless of how the shape is posed in space.

Geodesic Metrics: Walking the Surface

The most direct way to achieve isometry invariance is to replace straight-line Euclidean distance with geodesic distance—the shortest path a microscopic ant would take walking along the object's surface. As Stefanie Wuhrer et al. demonstrate in Posture invariant gender classification for 3D human models, measuring surface-bound distances allows for accurate classification of physical traits without requiring complex pose normalization or skeletal rigging.

This preservation of geodesic distance serves as a powerful constraint. For instance, Ajad Chhatkuli et al. in Model-free Consensus Maximization for Non-Rigid Shapes use geodesic consistency to filter out incorrect point correspondences across different subjects. Similarly, Eduard Serradell et al. show in Robust non-rigid registration of 2D and 3D graphs that geodesic paths along branching structures like blood vessels act as robust invariants during non-linear deformations.

Modern deep learning architectures have embraced this concept by embedding geodesic logic directly into neural operations. Jonathan Masci et al. introduced Geodesic Convolutional Neural Networks (GCNNs) in Geodesic Convolutional Neural Networks on Riemannian Manifolds, defining local convolution patches via geodesic polar coordinates to make the network inherently invariant to non-rigid deformations. To bypass non-differentiable classical algorithms, Puhua Jiang et al. propose Neural Intrinsic Embedding for Non-rigid Point Cloud Matching, which trains a network to map point cloud vertices into a high-dimensional space where Euclidean distances directly mirror the surface's intrinsic geodesic distances.

Inner Distances and Volumetric Thickness

While surface geodesics work well for thin shells, many real-world objects—like flexible proteins or 2D silhouettes—are better analyzed by looking inside their boundaries. This is the core idea behind the Inner Distance. Instead of measuring a straight line through external space or sticking strictly to the outer surface, the inner distance computes the shortest path restricted entirely to the interior of the object's volume or silhouette.

In IDSS: deformation invariant signatures for molecular shape comparison, Yu-Shen Liu et al. demonstrate that this volumetric shortest path metric remains highly stable across articulated deformations, making it ideal for matching flexible molecules in different conformational states. This approach was later extended into a web-based retrieval system by Yu-Shen Liu et al. in 3DMolNavi: A web-based retrieval and navigation tool for flexible molecular shape comparison. For 2D point set registration, Gang Wang 0008 et al. utilize inner-distance shape descriptors in Context-Aware Gaussian Fields for Non-rigid Point Set Registration to maintain alignment invariance when parts of an object (like a pair of scissors) articulate.

A closely related intrinsic property is local thickness or local diameter. By casting rays from a surface point through the internal volume to the opposing surface, algorithms can measure the local "bulk" of an object. Yi Fang et al. show in Three dimensional shape comparison of flexible proteins using the local-diameter descriptor that local thickness measurements are remarkably resilient to large-scale hinge movements because the internal volume of a component is largely unaffected by how it bends relative to another component.

Skeletal and Spectral Representations

Another powerful way to capture intrinsic structure is to compress the shape into its topological backbone: the skeleton or Medial Axis Transform (MAT). The medial axis represents the locus of centers of bi-tangent inscribed spheres within the shape. Because it captures both the internal connectivity and local thickness simultaneously, it provides an elegant, deformation-invariant representation.

In A skeletal measure of 2D shape similarity, Andrea Torsello et al. show that the ratio of boundary length change to skeletal length remains invariant under bending deformations. This occurs because boundary loss on the concave side of a bend is mathematically balanced by the boundary gain on the convex side. To bring this structural stability into deep learning, Jianwei Hu 0003 et al. developed MAT-Net: Medial Axis Transform Network for 3D Object Recognition, which leverages medial spheres to recognize non-rigid objects across varied poses without relying on volatile surface coordinates.

Finally, when shapes undergo deformations that are close to isometric but include minor stretching, spectral and conformal geometry offer a solution. David Gu et al. argue in Curve Space: Classifying Curves On Surfaces that by decoupling the conformal factor (which captures stretching) from the mean curvature (which captures bending), models can isolate the underlying identity of a non-rigid surface from its temporary pose. For pure function transfer across manifolds, Lukas Koestler et al. introduce Intrinsic Neural Fields: Learning Functions on Manifolds, using the eigenspectrum of the Laplace-Beltrami Operator to create neural fields that naturally adapt when a shape bends or deforms.

Key Foundational Work

A skeletal measure of 2D shape similarity (2004) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_isometric_deformation_invariance_ec78123d,
  title = {isometric deformation invariance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/isometric-deformation-invariance/rdr_ec78123ddd2d404d88d6c0987518e93f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_isometric_deformation_invariance_ec78123d_v1,
  title = {isometric deformation invariance},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/isometric-deformation-invariance/rdr_ec78123ddd2d404d88d6c0987518e93f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake