MASS: Multi-Attentional Semantic Segmentation of LiDAR Data for Dense Top-View Understanding

At the heart of all automated driving systems is ability to sense surroundings, e.g., through semantic segmentation LiDAR sequences, which experienced a remarkable progress due release large datasets such as SemanticKITTI and nuScenes-LidarSeg. While most previous works focus on xmlns:xlink="http://www.w3.org/1999/xlink">sparse input, xmlns:xlink="http://www.w3.org/1999/xlink">dense output masks provide self-driving cars with almost complete environment information. In this paper, we introduce MASS - Multi-Attentional Semantic Segmentation model specifically built for dense top-view understanding scenes. Our framework operates pillar- occupancy features comprises three attention-based building blocks: (1) keypoint-driven graph attention, (2) an LSTM-based attention computed from vector embedding spatial (3) pillar-based resulting in 360° mask. With extensive experiments both, nuScenes-LidarSeg, quantitatively demonstrate effectiveness our model, outperforming state art by 19.0% reaching 30.4% mIoU where first work addressing task. Furthermore, multi-attention shown be very effective 3D object detection validated KITTI-3D dataset, showcasing its high generalizability other tasks related vision.