Semantic segmentation for robotic systems can enable a wide range of applications, from self-driving cars to augmented reality systems. LiDAR scans can have various different characteristics and properties, such as number of beams, vertical FoV, angular resolution. Our method provides a framework to seamlessly train on pointcloud data from various different LiDAR sensor models and types.

S2AE was written using PyTorch (https://pytorch.org/) and depends on a few libraries.

• s2cnn: https://github.com/jonas-koehler/s2cnn
• lie_learn: https://github.com/AMLab-Amsterdam/lie_learn
• pynvrtc: https://github.com/NVIDIA/pynvrtc

Submodule references to these repositories can be found in the deps folder

Our paper is available at

Bernreiter, Lukas, Lionel Ott, Roland Siegwart, and Cesar Cadena. "SphNet: A Spherical Network for Semantic Pointcloud Segmentation" [ArXiv]