Efficient Convolutions for Real-Time Semantic Segmentation of 3D Point Clouds

In this work, we propose a novel voxel representation which allows for efficient, real-time processing of point clouds with deep neural networks. Our approach takes a 2D representation of a simple occupancy grid and produces fine-grained 3D segmentation. We show that our approach outperforms the state-of-the art while being an order of magnitude faster. We can perform segmentation of large outdoor scenes of size 160m x 80m in as little as 30ms. In indoor scenarios, we can segment full rooms in less than 15ms. This is crucial for robotics applications which require real-time inference for safety critical tasks.