Multi-Scale Voxel Class Balanced ASPP for LIDAR Pointcloud Semantic Segmentation

This paper explores efficient techniques to improve PolarNet model performance to address the real-time semantic segmentation of LiDAR point clouds. The core framework consists of an encoder network, Atrous spatial pyramid pooling (ASPP)/Dense Atrous spatial pyramid pooling (DenseASPP) followed by a decoder network. Encoder extracts multi-scale voxel information in a top-down manner while decoder fuses multiple feature maps from various scales in a bottom-up manner. In between encoder and decoder block, an ASPP/DenseASPP block is inserted to enlarge receptive fields in a very dense manner. In contrast to PolarNet model, we use weighted cross entropy in conjunction with Lovasz-softmax loss to improve segmentation accuracy. Also this paper accelerates training mechanism of PolarNet model by incorporating learning-rate schedulers in conjunction with Adam optimizer for faster convergence with fewer epochs without degrading accuracy. Extensive experiments conducted on challenging SemanticKITTI dataset shows that our high-resolution-grid model obtains competitive state-of-art result of 60.6 mIOU @21fps whereas our low-resolution-grid model obtains 54.01 mIOU @35fps thereby balancing accuracy/speed trade-off.