MeshCNN: A Network with an Edge for 3D Shape Analysis

Explore a cutting-edge approach to 3D shape analysis in this 46-minute conference talk by Shahar Fleishman at BIMSA. Delve into MeshCNN, a convolutional neural network specifically designed for triangular meshes. Learn how this innovative network leverages the unique properties of polygonal meshes to directly analyze 3D shapes. Discover how MeshCNN combines specialized convolution and pooling layers that operate on mesh edges, utilizing their intrinsic geodesic connections. Understand the task-driven pooling process that allows the network to expose and expand important features while discarding redundant ones. Gain insights into the effectiveness of this approach across various learning tasks applied to 3D meshes, and see how it overcomes the challenges posed by the non-uniformity and irregularity of mesh representations in traditional neural network analyses.

Shahar Fleishman: Meshcnn a network with an edge #ICBS2024