Abstract:We present GeGnn, a learning-based method for computing the approximate geodesic distance between two arbitrary points on discrete polyhedra surfaces with constant time complexity after fast precomputation. Previous relevant methods either focus on computing the geodesic distance between a single source and all destinations, which has linear complexity at least or require a long precomputation time. Our key idea is to train a graph neural network to embed an input mesh into a high-dimensional embedding space and compute the geodesic distance between a pair of points using the corresponding embedding vectors and a lightweight decoding function. To facilitate the learning of the embedding, we propose novel graph convolution and graph pooling modules that incorporate local geodesic information and are verified to be much more effective than previous designs. After training, our method requires only one forward pass of the network per mesh as precomputation. Then, we can compute the geodesic distance between a pair of points using our decoding function, which requires only several matrix multiplications and can be massively parallelized on GPUs. We verify the efficiency and effectiveness of our method on ShapeNet and demonstrate that our method is faster than existing methods by orders of magnitude while achieving comparable or better accuracy. Additionally, our method exhibits robustness on noisy and incomplete meshes and strong generalization ability on out-of-distribution meshes. The code and pretrained model can be found on https://github.com/IntelligentGeometry/GeGnn.

Using Locally Estimated Geodesic Distance to Optimize Neighborhood Graph for Isometric Data Embedding

Locality Pursuit Embedding

Improving Geodesic Distance Estimation Based on Locally Linear Assumption

Image Feature Optimization Based on Nonlinear Dimensionality Reduction

GeodesicEmbedding (GE): A High-Dimensional Embedding Approach for Fast Geodesic Distance Queries

Estimating Geodesic Distances on Locally Linear Patches

Deep Manifold Computing and Visualization Using Elastic Locally Isometric Smoothness

Isometric Immersion Learning with Riemannian Geometry

A Regularized Approach for Geodesic-Based Semisupervised Multimanifold Learning

Uniform Interpolation Constrained Geodesic Learning on Data Manifold

Geodesic Learning with Uniform Interpolation on Data Manifold

Parallel Vector Field Embedding.

Extended ISOMAP Based on Neighborhood Sets Relation

Global structure-guided neighborhood preserving embedding for dimensionality reduction

Learning the Geodesic Embedding with Graph Neural Networks

Short and Straight: Geodesics on Differentiable Manifolds

Isomap Based on the Image Euclidean Distance

Adaptive Neighboring Selection Algorithm Based on Curvature Prediction in Manifold Learning

Normed Spaces for Graph Embedding

DIMAL: Deep Isometric Manifold Learning Using Sparse Geodesic Sampling

Neighborhood preserving embedding on Grassmann manifold for image-set analysis