📊 statistics

A Functional Approach to Curve Alignment and Shape Analysis

This paper introduces a novel Functional Data Analysis framework that utilizes basis expansion techniques to derive analytic solutions for curve alignment and develop a generative model for random contours, thereby overcoming the limitations of discrete observation methods in capturing continuous geometric structures and deformation dependencies.

Original authors: Issam-Ali Moindjié, Cédric Beaulac, Marie-Hélène Descary

Published 2026-05-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Issam-Ali Moindjié, Cédric Beaulac, Marie-Hélène Descary

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a computer to recognize different shapes, like a heart, a butterfly, or a fork. You show it a bunch of pictures. But there's a problem: in some pictures, the heart is upside down; in others, it's tiny; in some, it's shifted to the left; and in others, the computer starts tracing the outline from the top point, while in another, it starts from the bottom.

If you just average these pictures together without fixing these differences, you don't get a clear heart. You get a blurry, unrecognizable blob. This is the problem the authors of this paper are solving. They call it "Curve Alignment."

Here is a simple breakdown of their solution, using everyday analogies:

1. The Problem: The "Messy Dance Floor"

Think of a group of dancers (the shapes) trying to perform the same routine.

Translation: Some dancers are standing in the corner; others are in the center.
Scaling: Some are doing the routine as giants; others as tiny people.
Rotation: Some are facing north; others are spinning around.
Reparametrization (The Tricky One): This is like the dancers starting the routine at different beats. One starts with a jump, another with a spin. Even though they are doing the same dance, their "start time" is different.

In the past, scientists tried to analyze these shapes by looking at them as a list of dots (pixels). The authors say this is like trying to understand a song by looking at a list of individual notes without hearing the melody. It misses the continuous flow of the shape.

2. The Solution: The "Smooth Tape" Approach

Instead of looking at dots, the authors treat the shape as a smooth, continuous line (like a piece of string). They use a mathematical tool called Functional Data Analysis (FDA).

Think of it like this: Instead of taking a photo of a dancer and counting their pixels, they record the dancer's movement on a smooth video tape. This allows them to handle shapes from different camera resolutions (high-def vs. low-def) without losing quality.

3. The "ICF Algorithm": The Magic Matchmaker

The core of their paper is a new method called the Iterative Closest Function (ICF) algorithm.

Imagine you have a "Reference Dance" (a template) and a "Messy Dance" (a new shape you want to align).

Step 1: Move and Resize. First, the algorithm moves the messy dancer to the center of the room and makes them the same size as the reference. This is easy.
Step 2: Spin and Sync. Now, the messy dancer is the right size but might be facing the wrong way or starting the dance at the wrong beat.
- The algorithm tries spinning the dancer until they face the right direction.
- Then, it tries shifting the "start time" of the dance (sliding the tape forward or backward) until the moves match the reference perfectly.

The authors found that for shapes coming from images (like silhouettes), the biggest confusion usually comes from where the computer decided to start drawing the line. Their method is specifically designed to find that "start point" and slide the line until it matches the template perfectly, without distorting the shape.

4. The "Two-Step" Analysis

Once the shapes are aligned (all facing the same way, same size, starting at the same beat), the authors do something clever. They don't just throw away the information about how the shapes were moved.

They split the data into two separate stories:

The Shape Story: What does the object actually look like? (e.g., Are the butterfly's wings wide or narrow?)
The Deformation Story: How was the object moved? (e.g., Was it rotated? Was it shifted?)

They use a statistical tool called PCA (Principal Component Analysis) on both stories separately.

Analogy: Imagine analyzing a choir.
- Story 1 (Shape): You analyze the different types of voices (soprano, bass) to see what makes a "good choir."
- Story 2 (Deformation): You analyze how much the singers swayed or moved around the stage.
- By keeping these separate, you can understand the choir's music and their stage presence without them confusing each other.

5. The Results: "Real" vs. "Fake" Shapes

The authors tested their method on two things:

Fake Data: They created computer-generated hearts with random rotations and starting points. Their method was much better at finding the "true" heart shape than older methods, which often got confused and produced distorted, "alien" shapes.
Real Data: They used pictures of butterflies and forks from a public database.
- Butterflies: Their model correctly identified that the main difference between butterflies was the size and shape of their wings.
- Forks: Their model noticed that some forks had three prongs and others had four.
- Comparison: When they tried to generate new shapes using their method, the results looked like real butterflies and forks. When they used older methods (which didn't align the shapes first), the generated "forks" looked like melted blobs or strange, unrecognizable objects.

The Bottom Line

This paper introduces a smarter way to teach computers to recognize shapes. Instead of treating every shape as a messy pile of dots, they treat them as smooth, continuous lines. They built a tool to automatically "straighten out" the lines so they all face the same way and start at the same time.

Once the lines are straight, the computer can clearly see the differences between a butterfly and a fork, and it can even generate new, realistic-looking shapes that look like the real thing. It's like taking a messy pile of tangled headphones, carefully untangling them, and then realizing they are all the same brand of headphones, just wrapped differently.

Technical Summary: A Functional Approach to Curve Alignment and Shape Analysis

Problem Statement
The paper addresses the statistical analysis of object shapes represented as contours extracted from images. While traditional shape analysis often relies on discrete observations (landmarks or point clouds), this approach overlooks the continuous nature of contours and their underlying geometric structure. Furthermore, contours extracted from images are subject to deformation variables including scaling ( $\rho$ ), translation ( $T$ ), rotation ( $O$ ), and reparametrization ( $\gamma$ ). A specific challenge in image-derived contours is phase variability caused by arbitrary starting points determined by edge detection algorithms. Existing methods, such as elastic shape analysis frameworks (e.g., Srivastava et al., 2010) or iterative closest point (ICP) algorithms, often rely on discretization, lack closed-form solutions, or fail to explicitly disentangle deformation parameters from the intrinsic shape, leading to potential loss of statistical information and reduced interpretability.

Methodology
The authors propose a framework grounded in Functional Data Analysis (FDA) that treats contours as bivariate functional variables $C(t) = (X(t), Y(t))^\top$ defined on $t \in [0, 1]$ . The core methodology consists of two main stages:

Curve Alignment via the Iterative Closest Function (ICF) Algorithm:
- Model: The observed curve is modeled as $C(t) = \rho O_\theta \tilde{C} \circ \gamma_\delta(t) + T$ , where $\tilde{C}$ is the latent shape (pre-shape).
- Basis Expansion: To handle the continuous nature of the data, curves are smoothed and represented using a Fourier basis expansion. This choice is motivated by the periodicity of closed curves and the specific time-shift property of Fourier functions, which simplifies the estimation of reparametrization.
- Reparametrization Space: The authors restrict the reparametrization function $\gamma$ to a specific space $\Gamma$ representing cyclic shifts of the starting point ( $\gamma_\delta(t) = \text{mod}(t-\delta, 1)$ ). This addresses the dominant source of misalignment in image contours without the complexity of full diffeomorphic warping.
- Estimation: The algorithm estimates deformation variables ( $\rho, T, \theta, \delta$ $ρ, T, θ, δ$ ) by minimizing the $L^2$ $L^{2}$ distance between a curve and a template (the Fréchet mean).
  - Translation and scaling are derived directly from the Fourier coefficients.
  - Rotation ( $\theta$ ) and reparametrization ( $\delta$ ) are estimated via an alternating optimization procedure (ICF). For a fixed $\delta$ , $\theta$ is solved via a Procrustes problem. For a fixed $\theta$ , $\delta$ is found by solving a trigonometric equation derived from the matrix representation of the basis functions.
- Fréchet Mean: An iterative procedure is employed to jointly estimate the Fréchet mean shape and align the curves, as the mean depends on the alignment and vice versa.
Generative Modeling via Separated PCA:
- Once aligned, the data is decomposed into a shape component and deformation components.
- Shape Modeling: The aligned pre-shapes (scaled to unit norm) are modeled using Multivariate Functional Principal Component Analysis (MFPCA) to capture intrinsic shape variability.
- Deformation Modeling: The deformation variables ( $\rho, \theta, \delta, T$ ) are transformed into an unconstrained Euclidean space ( $Z_2$ ) and modeled using standard Multivariate PCA.
- Generative Model: A joint generative model is constructed by assuming the concatenated score vectors follow a multivariate Gaussian distribution. This allows for the simulation of new curves that preserve the dependency between shape and deformation.

Key Contributions

Functional Alignment Procedure: The introduction of the ICF algorithm, which operates directly in the functional space using Fourier basis expansions. This avoids the need for additional discretization beyond the basis representation and yields analytic solutions for the estimation of deformation variables.
Explicit Deformation Modeling: Unlike approaches that treat deformations as nuisance parameters, this framework explicitly estimates and models scaling, translation, rotation, and starting-point shifts. This allows for the separate analysis of shape variability and deformation effects.
Separate PCA Framework: A novel generative model that performs separate PCA on the aligned shape and the deformation parameters. This preserves the interpretability of principal components, distinguishing between modes of variation driven by shape geometry versus those driven by alignment parameters.
Handling Phase Variability: The specific restriction of the reparametrization space to cyclic shifts addresses the practical issue of arbitrary starting points in image contours, providing a tractable solution where more flexible elastic frameworks may struggle with identifiability.

Results

Simulation Study: The ICF algorithm was compared against the Iterative Closest Point (ICP) algorithm and the Square-Root Velocity Function (SRVF) elastic alignment method.
- Accuracy: ICF demonstrated superior performance in estimating the reparametrization parameter $\delta$ and rotation $\theta$ , particularly at lower sampling resolutions. ICP suffered from large errors due to its operation on unordered point clouds, while SRVF struggled with rotation estimation, likely due to the partial compensation between rotation and reparametrization in elastic frameworks.
- Robustness: The functional approach proved robust to discretization levels.
Real Data Analysis (MPEG-7 Database): The method was applied to butterfly and fork object classes.
- Alignment: The procedure successfully aligned curves with varying orientations and starting points, producing coherent Fréchet mean shapes.
- Interpretability: The separated PCA revealed distinct modes of variation. For butterflies, the first component captured wing scaling, while subsequent components captured wing geometry. For forks, components distinguished between three-pronged and four-pronged structures.
- Generative Performance: Curves generated by the proposed model were more realistic and structurally consistent (e.g., preserving antennae in butterflies and fork tines) compared to a baseline MFPCA applied to unaligned curves, which produced distorted or incoherent shapes.

Significance and Claims
The paper claims that its functional approach offers a "simple, identifiable, and computationally efficient alternative" to complex elastic alignment frameworks, specifically tailored for contours automatically extracted from images. By focusing on the dominant source of phase variability (the starting point) and utilizing the properties of Fourier bases, the method achieves curve alignment without the computational burden of numerical optimization over diffeomorphisms.

The authors emphasize that their framework bridges Functional Data Analysis, shape analysis, and image analysis. A significant advantage highlighted is the ability to work within a linear functional space by explicitly modeling scaling, thereby avoiding the complexities of non-linear manifolds (such as the pre-shape sphere) required when treating scaling as a nuisance parameter. The work demonstrates that explicitly modeling deformation variables leads to better interpretability and more realistic generative models compared to methods that conflate shape and alignment variability. The authors acknowledge limitations, noting the current focus on single-object contours and the restriction to starting-point shifts, suggesting future work could extend to multiple interacting curves and more flexible reparametrization classes.