Abstract

Adapting the power of Deep Learning to other types of data than images is not straightforward. Instead of projecting the data into grids, so that standard CNN can process them, we chose to define a convolution that operates directly on 3D points. Inspired by the design of image convolutions, and by hand crafted features on point clouds, we created the Kernel Point Convolution (KPConv), and CNN architectures for classification and segmentation of 3D point clouds. Our networks outperform state of the art methods in almost any situation.

Organizational Details

How to design a 3D point convolution, the example of KPConv

Date: Monday, 9th December 2019
Time: 04:00 pm
Room: O27/331