TL;DR
Linear Spherical Sliced Optimal Transport (LSSOT) is a fast, novel framework that embeds spherical probability distributions into a flat $L^2$ space using hypersphere slicing techniques. By combining slicing with linear circular optimal transport, LSSOT reduces the computational cost of large-scale pairwise distribution comparisons from quadratic to linear. This enables massive speedups in demanding geometric tasks like brain cortical surface registration without sacrificing accuracy.
What Problem Does This Paper Address?
Spherical data is foundational to many fields, appearing in omnidirectional computer vision, global climate modeling, and medical imaging—such as analyzing the genus-zero topology of the human cerebral cortex. To compare these data distributions accurately, researchers rely on metrics that respect the sphere's non-Euclidean geometry.
Optimal Transport (OT) is the gold standard for this task because it accounts for underlying geometry, but it suffers from a prohibitive cubic computational complexity ($O(N^3)$ relative to sample size). Sliced OT (SOT) mitigates this by projecting spherical data onto 1D circles, but calculating pairwise distances across a large dataset of $K$ objects still scales quadratically ($O(K^2)$). This paper introduces a framework to scale spherical data comparison to large datasets by creating a geometrically faithful, linear embedding space.
Methods
The authors introduce Linear Spherical Sliced Optimal Transport (LSSOT). The core methodology consists of two main steps:
- Spherical Slicing: The hypersphere is sliced into 1D great circles using projections defined over the Stiefel manifold.
- Linear Embedding: Instead of computing standard OT on these circles, LSSOT applies Linear Circular Optimal Transport (LCOT) to embed each circular projection into a flat $L^2$ space relative to a fixed, uniform reference measure.
To handle the numerical singularity where projections at the poles are ill-defined, the authors implement an $\epsilon$-cap around the poles, evenly redistributing polar mass to the surrounding discrete points. Crucially, the authors provide a rigorous mathematical proof that LSSOT is a true metric for continuous probability density functions, resolving a common theoretical limitation of sliced transport methods.
Datasets
The framework was validated across several diverse datasets:
- NKI Dataset: 515 subjects with T1-weighted brain MRI scans, processed into spherical cortical surface meshes (40,962 vertices) to evaluate sulcal depth and parcellations.
- ADNI Dataset: 412 subjects processed similarly to the NKI dataset for atlas-based cortical registration.
- ModelNet: A standard 3D CAD dataset used to evaluate 3D point cloud interpolation via gradient flow.
- SHREC 2007 (Watertight Track): Mesh models from 4 genus-0 closed surface categories (human, octopus, teddy bear, fish) used for 3D shape clustering and retrieval.
Key Insights and Ideas
- The Power of a Fixed Template: In atlas-based applications like cortical registration, the reference template's embedding is computed once before training. During training iterations, only the moving distribution needs to be embedded, turning a complex pairwise transport optimization into a simple $L^2$ distance check.
- Linear Scaling: By embedding distributions into a unified $L^2$ space, pairwise distance matrix calculations scale as $O(K)$ instead of $O(K^2)$ for the transport step, unlocking massive scalability for group analysis.
- Theoretical Rigor: The paper proves that both LSSOT and the baseline Spherical Sliced Wasserstein (SSW) discrepancy are true metrics (not just pseudo-metrics) when restricted to continuous densities, filling a significant gap in the theoretical literature.
Key Results
- Massive Speedups: In pairwise distance computations, LSSOT dramatically outperforms methods like Sinkhorn, SSW, and standard OT, maintaining near-flat runtime scaling as the number of distributions increases.
- Cortical Surface Registration: When integrated into a neural network framework (S3Reg) on the NKI dataset, LSSOT achieved registration accuracy (Dice score of 0.846) on par with the state-of-the-art SSW method (0.842) but reduced training time by over 90% (101 seconds per epoch vs. 1350 seconds for SSW).
- Geometric Fidelity: LSSOT successfully preserves linear transitions in scaling and rotation geometries, as demonstrated through Von Mises-Fisher (VMF) distribution toy experiments and smooth point cloud interpolations.
Limitations and Future Work
- Probability Measures Only: LSSOT is strictly defined for probability distributions, meaning the input data must be normalized and have equal total mass. This can lead to information loss for certain types of raw physical signals.
- Future Directions: The authors plan to extend LSSOT to support positive signals with unequal mass (partial optimal transport) and non-positive spherical signals, starting on $S^1$ before scaling back up to $S^2$ via slicing.
Summary updated