Parallelised Differentiable Straightest Geodesics for 3D Meshes

A research team from Imperial College London and the Technical University of Munich has released 'digeo', a breakthrough computational library that solves a fundamental mathematical bottleneck in 3D machine learning. The library implements parallelised differentiable straightest geodesics—essentially providing a way to compute the shortest paths on curved 3D surfaces (meshes) in a way that's compatible with gradient-based learning. This gives researchers a differentiable exponential map, a core operator from Riemannian geometry that was previously unavailable in closed form for discrete surfaces. The implementation achieves 10-100x speedups through GPU parallelization compared to traditional CPU methods.

The technical advance enables three immediate applications: a new geodesic convolutional layer for neural networks operating directly on mesh surfaces, a flow matching method for generative modeling of 3D shapes, and a second-order optimizer applied to centroidal Voronoi tessellation. The library addresses what the authors identify as the main obstacles to machine learning on meshes: lack of closed-form Riemannian operators, non-differentiability of discrete counterparts, and poor parallelization. By making these geometric computations both differentiable and fast, 'digeo' bridges the gap between theoretical Riemannian geometry and practical deep learning pipelines for 3D data.

Accepted to CVPR 2026, the work represents a significant step toward geometrically accurate machine learning on non-Euclidean domains. The researchers provide both an extrinsic proxy function approach and a geodesic finite differences scheme for differentiation, along with comprehensive accuracy proofs. The pip-installable library lowers the barrier to entry for researchers working on 3D computer vision, graphics, and geometric deep learning, potentially accelerating development in areas like digital humans, medical imaging, and physics simulation.