3D-MiniNet: Learning a 2D Representation from Point Clouds for Fast and Efficient 3D LIDAR Semantic Segmentation

LIDAR semantic segmentation, which assigns a semantic label to each 3D point\nmeasured by the LIDAR, is becoming an essential task for many robotic\napplications such as autonomous driving. Fast and efficient semantic\nsegmentation methods are needed to match the strong computational and temporal\nrestrictions of many of these real-world applications.\n This work presents 3D-MiniNet, a novel approach for LIDAR semantic\nsegmentation that combines 3D and 2D learning layers. It first learns a 2D\nrepresentation from the raw points through a novel projection which extracts\nlocal and global information from the 3D data. This representation is fed to an\nefficient 2D Fully Convolutional Neural Network (FCNN) that produces a 2D\nsemantic segmentation. These 2D semantic labels are re-projected back to the 3D\nspace and enhanced through a post-processing module. The main novelty in our\nstrategy relies on the projection learning module. Our detailed ablation study\nshows how each component contributes to the final performance of 3D-MiniNet. We\nvalidate our approach on well known public benchmarks (SemanticKITTI and\nKITTI), where 3D-MiniNet gets state-of-the-art results while being faster and\nmore parameter-efficient than previous methods.\n