Numerical Inversion of SRNF Maps for Elastic Shape Analysis of Genus-Zero Surfaces

Hamid Laga, Qian Xie, Ian H. Jermyn, Anuj Srivastava

Abstract

Recent developments in elastic shape analysis (ESA) provide a comprehensive framework for simultaneous registration, deformation, and comparison of shapes. These methods achieve computational efficiency using square-root representations that transform invariant elastic metrics into Euclidean metrics. For analyzing shapes of embeddings in $\mathbb{R}^3$, Square-Root Normal Fields (SRNFs) transform an elastic metric into the $L^2$ metric. However, a critical need in shape analysis is a method for inverting solutions computed in SRNF space back to the original surface space for visualization and inference. This paper presents a numerical approach using an efficient multiresolution algorithm to estimate surfaces corresponding to given SRNFs. The solution is found to be effective for complex shapes undergoing significant deformations, such as human bodies and brain structures.

Introduction: The Inversion Challenge

In 3D computer vision and medical imaging, comparing shapes requires a framework that accounts for "elastic" changes—bending and stretching—rather than just rigid transformations. Elastic Shape Analysis (ESA) provides this, but the underlying mathematics is computationally demanding. To simplify these operations, researchers utilize Square-Root Normal Fields (SRNFs). SRNFs map complex surfaces into a Hilbert space where the "distance" between shapes is a standard Euclidean ($L^2$) distance.

While the mapping from a surface to an SRNF is straightforward, the inverse mapping is mathematically "hard." When researchers compute a mean shape or a deformation path in the SRNF space, the result is a mathematical object that does not immediately correspond to a physical surface. Without a robust inversion method, these results cannot be visualized or used for practical inference in 3D space. This paper introduces a numerical "undo" button for this process.

Methodology: Multiresolution Optimization

The authors frame SRNF inversion as a non-linear optimization problem. Given a target SRNF $q$, the goal is to find a surface $f$ such that its SRNF, $Q(f)$, is as close to $q$ as possible.

A naive single-step optimization often fails, getting trapped in local minima or producing jagged, non-physical results. To overcome this, the paper introduces a multiresolution approach. The algorithm begins with a coarse version of the shape and progressively refines it across multiple scales.

Figure 5: The multiresolution approach (a) progressively refines the surface from a coarse initialization, avoiding the artifacts seen in a single-step inversion (bottom).

Key innovations supporting this process include:

  • Basis Functions: Surfaces are represented using Spherical Harmonics (SH) for general shapes or a Principal Component Analysis (PCA) basis when training data is available. The PCA basis is significantly more efficient, requiring approximately 100 elements compared to over 3,000 for SH.
  • Analytical Initialization: For star-shaped surfaces, a closed-form analytical inversion exists. This provides a high-speed starting point for the optimization of more complex geometries.
  • Joint Registration: The inversion process handles "registration" (aligning points between shapes) by incorporating rotations and re-parameterizations directly into the optimization loop.

Results and Applications

The proposed method was validated across diverse datasets, demonstrating its versatility and accuracy.

Reconstruction and Speed

On the SHREC07 dataset of 400 models, the median reconstruction error was remarkably low (0.0008 relative to unit surface area), indicating a nearly lossless inversion. Furthermore, the use of a PCA basis reduced inversion time from 16 minutes (using SH) to just 20 seconds, making large-scale statistical analysis feasible.

Human Body and Medical Imaging

The method excels at handling articulated shapes like human bodies. By performing statistical operations in the SRNF space and inverting the results, the authors can synthesize natural-looking deformations and "average" poses without the shrinkage artifacts common in linear interpolation.

Figure 6: The inversion method successfully reconstructs complex, articulated genus-zero surfaces like the human body.

In a clinical application involving brain subcortical structures, the SRNF-based framework was used to classify ADHD. The resulting classifier achieved 94.1% accuracy for the left putamen, significantly outperforming traditional methods like SPHARM (50.0%) and Iterative Closest Point (61.8%).

Conclusion

The introduction of a robust numerical inversion for SRNFs bridges the gap between abstract shape mathematics and practical 3D visualization. By leveraging multiresolution optimization and efficient basis representations, the authors have enabled high-fidelity elastic shape analysis for complex genus-zero surfaces. While currently limited to sphere-equivalent topologies, this work sets the stage for future extensions to higher-genus surfaces and more sophisticated generative models in the SRNF space.

Summary updated

Abstract

Recent developments in elastic shape analysis (ESA) are motivated by the fact that it provides a comprehensive framework for simultaneous registration, deformation, and comparison of shapes. These methods achieve computational efficiency using certain square-root representations that transform invariant elastic metrics into euclidean metrics, allowing for the application of standard algorithms and statistical tools. For analyzing shapes of embeddings of in , Jermyn et al. [1] introduced square-root normal fields (SRNFs), which transform an elastic metric, with desirable invariant properties, into the metric. These SRNFs are essentially surface normals scaled by square-roots of infinitesimal area elements. A critical need in shape analysis is a method for inverting solutions (deformations, averages, modes of variations, etc.) computed in SRNF space, back to the original surface space for visualizations and inferences. Due to the lack of theory for understanding SRNF maps and their inverses, we take a numerical approach, and derive an efficient multiresolution algorithm, based on solving an optimization problem in the surface space, that estimates surfaces corresponding to given SRNFs. This solution is found to be effective even for complex shapes that undergo significant deformations including bending and stretching, e.g., human bodies and animals. We use this inversion for computing elastic shape deformations, transferring deformations, summarizing shapes, and for finding modes of variability in a given collection, while simultaneously registering the surfaces. We demonstrate the proposed algorithms using a statistical analysis of human body shapes, classification of generic surfaces, and analysis of brain structures.
BibTeX
@article{laga2017numerical,
  author = {Hamid Laga and Qian Xie and Ian H. Jermyn and Anuj Srivastava},
  title = {{Numerical Inversion of SRNF Maps for Elastic Shape Analysis of Genus-Zero Surfaces}},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year = {2017},
  doi = {10.1109/tpami.2016.2647596},
}
Figure