EfficientPillarNet: A Fast Deep Network for Semantic Segmentation of Large-scale Point Clouds from Lidar

Semantic segmentation in large-scale point clouds has brought increasing research in the 3D vision inspection field of many robotic applications, such as auto-driving and drone. To segment 3D point clouds end-to-end, there are two typical ways recently: using 3D convolution on the structure-specific 3D point clouds tends to be more accurate, while encoding point clouds into 2D pseudo graph tend to be faster. In this paper, we propose EfficientPillarNet, a real-time processing network, which projects disorder point clouds into bird's-eye-view by using factorized convolutions and dilated convolutions in order to gain state-of-art operating efficiency while remaining outstanding performance. Our approach can realize pixel-wise semantic segmentation in real-time in a single GPU. Our experiment shows that our model's segmentation performance can run at 20Hz on the 1080Ti and main classes of mIoU can reach 58.8% on the lidar dataset. This makes our model an ideal approach for scene understanding of large-scale point clouds.