Finding correspondences between two 3D objects—identifying which point on a scanned human's hand matches the hand of a different scan—is a fundamental task in computer vision. For rigid objects, this is a matter of simple rotation and translation. However, real-world objects like humans, animals, and cloth are non-rigid. They bend, stretch, and twist.
When a person moves from a standing to a sitting position, the Euclidean (straight-line) distance between their hand and their knee changes drastically. Traditional matching algorithms that rely on these external spatial coordinates fail because the "shape" of the object in 3D space is constantly shifting. To solve this, researchers have moved toward intrinsic geometry, treating shapes not as collections of points in a 3D box, but as manifolds where the only distances that matter are those measured along the surface itself.
The Ant on the Surface: Geodesic Invariance
The core intuition behind non-rigid matching is the "ant's eye view." If an ant walks along the surface of a human body from the elbow to the wrist, that distance (the geodesic distance) remains roughly the same whether the arm is bent or straight. As Mickaël Péchaud and Laurent Cohen explain in Geodesic Methods in Computer Vision and Graphics, these intrinsic metrics are invariant to isometric (bending) transformations.
To exploit this, early methods like those proposed by Stefanie Wuhrer in Posture invariant surface description and feature extraction attempted to "unfold" or "flatten" 3D shapes into a canonical form using Multi-Dimensional Scaling (MDS). By mapping the complex, bending 3D mesh into a lower-dimensional space where geodesic distances become Euclidean distances, the problem of non-rigid matching is transformed back into a simpler rigid alignment problem.
From Points to Functions: Spectral Matching
Even with geodesic distances, matching thousands of individual points is a combinatorial nightmare—an NP-hard problem. A major breakthrough occurred by shifting the perspective from matching points to matching functions defined on the surface.
Every shape has a unique "vibrational signature" defined by the eigenfunctions of its Laplace-Beltrami operator (similar to how a guitar string has fundamental frequencies and harmonics). In Spectral Generalized Multi-Dimensional Scaling, Anastasia Dubrovina and Ron Kimmel showed that we can represent the correspondence between two shapes as a small matrix—a Functional Map.
Instead of asking "Where does point A go?", we ask "How does a function (like a heat map) on Shape 1 transfer to Shape 2?" This representation, further refined into deep learning frameworks like FMNet by Or Litany and Emanuele Rodolà in Deep Functional Maps, allows for end-to-end training. The network learns descriptors that are specifically optimized to produce clean, low-rank functional maps, effectively ignoring the "noise" of non-rigid deformation.
Bridging the Gap with Intermediate States
A persistent challenge in non-rigid matching is the "large deformation" problem: matching a person in a fetal position to a person doing a jumping jack. The geometric gap is often too wide for direct descriptors to bridge.
Recent work suggests that the solution lies in the "neighborhood" of shapes. Aysim Toker and Daniel Cremers introduced G-MSM (Unsupervised Multi-Shape Matching with Graph-based Affinity Priors), which propagates correspondence information through a sequence of similar intermediate shapes. If Shape A is too different from Shape C, the algorithm finds a Shape B that acts as a bridge.
This concept of "transitive consistency" is also central to GenCorres by Haitao Yang and Qixing Huang. By using generative models to "fill in" the space between poses, these methods ensure that correspondences are cycle-consistent across an entire collection, making the matching process robust even when direct pairwise comparisons fail.
Handling Non-Isometric Realities
While many models assume "perfect" bending (isometry), real-world data involves stretching, noise, and missing parts. Mahdi Saleh and colleagues addressed this in Bending Graphs by using Gated Optimal Transport (GOT). This mechanism allows the model to "gate" or ignore unreliable matches caused by extreme stretching or occlusions, focusing only on the most confident structural correspondences. Similarly, Abed Malti integrated physics-based mechanical priors in A linear least-squares solution to elastic Shape-from-Template, ensuring that the predicted deformations follow the laws of elasticity rather than just geometric heuristics.