Lightningnet: Fast and Accurate Semantic Segmentation for Autonomous Driving Based on 3D LIDAR Point Cloud

Semantic segmentation is A series of earlier works have demonstrated that 3D LiDAR point cloud segmentation is a promising approach. However, these methods usually rely on pretrained models. And their requiring improvements in either speed or accuracy prevent them to be applied to mobile platforms. To address these problems, a smaller Convolutional Neural Network (CNN) architecture, namely, LightningNet is presented. Furthermore, we propose a lightweight pipeline with Feature Refinement Modules to boost the performance of real-time 3D LiDAR point cloud segmentation. Our model is trained on spherical images projected from LiDAR point clouds on KITTI [1] dataset. Experiments show accuracy improvements of 4.1-8.2% in small objects and 4% in terms of mIoU over the state-of-the-art of spherical-image based method with 111 fps on a single 1080Ti GPU and 14fps on a Jetson TX2, which enables deployment for real-time semantic segmentation in autonomous driving. We achieve superior performance on another large dataset called SemanticKITTI (illustrated in Figure 1) both in speed and accuracy also.