Range-view (RV) Based Semantic Segmentation of Outdoor Point Cloud with Data Augmentation

Abstract

Semantic segmentation is a critical task in scene understanding of autonomous driving. Because of the irregular and sparse structure of outdoor point clouds, semantic segmentation for point clouds has received a lot of attention from both academia and industry. In this paper, we propose an approach that combines data augmentation for point clouds and lightweight 2D semantic segmentation network. The data augmentation technique produces a balanced dataset for training by interpolating more samples of object classes. The 2D deep layer aggregation network is then employed to train a semantic segmentation model on above augmented dataset to achieve better performance while costing less memory. We benchmark our model on the NuScenes dataset against RangeNet++. Our experiments demonstrate a +1.8% rise in mIoU and an over 6-fold reduction in trainable parameters compared to the baseline.