SVASeg: Sparse Voxel-Based Attention for 3D LiDAR Point Cloud Semantic Segmentation

3D LiDAR has become an indispensable sensor in autonomous driving vehicles. In LiDAR-based point cloud semantic segmentation, most voxel-based segmentors cannot efficiently capture large amounts of context information, resulting limited receptive fields and limiting their performance. To address this problem, a sparse attention network is introduced for termed SVASeg, which captures information between voxels through multi-head (SMHA). The traditional directly be applied to the non-empty voxels. end, hash table built according incrementation voxel coordinates lookup neighboring each voxel. Then, are grouped into different groups, group corresponds local region. Afterwards, position embedding, feature fusion performed aggregate information. Based on SMHA module, SVASeg can operate voxels, maintaining comparable computational overhead convolutional method. Extensive experimental results SemanticKITTI nuScenes datasets show superiority SVASeg.