Le Clerc,F., Sun,H.
Abstract:
Processing 3D meshes using convolutional neural networks requires convolutions to operate on features sampled on non-Euclidean manifolds. To this purpose, spatial-domain approaches applicable to meshes with different topologies locally map feature values in vertex neighborhoods to "patches" that define the inputs to the convolution filters in a consistent manner for each vertex. We show that this generalization of the convolution operator significantly increases the memory footprint of convolutional layers and sets a practical limit to network depths. We propose a memory-optimized convolution scheme that mitigates the issue and allows more convolutional layers to be included in a network for the same memory budget. We demonstrate the benefits of our approach on mesh registration experiments and show that deeper networks bring substantial improvements to the registration accuracy. Our results outperform the state of art.