Abstract: An unsupervised domain adaptation for LiDAR semantic segmentation method is proposed to deal with the problem of excessive data required for semantic segmentation network training in outdoor large-scale scenes. The method uses a modified RandLA-Net for semantic segmentation using a small number of point clouds from the SPTLS3D’s real world data as target objects. The model finishes the pre-training of the segmentation network on SensatUrban， and completes the transfer task by minimizing the domain gap between the source and target domains. The RandLA-Net losses the global features of the original point cloud in the encoding process， so an additional method of obtaining global information to join the network decoding is proposed. In addition， for getting the differentiated information， the weights of the local attention module of RandLA-Net is changed to use the difference between the features of each point and the average features of its neighbors. The experiments show that the mean intersection over union of the network are 54.3% on SemanticKITTI and 71.91% on Semantic3D. The mIoU of the pre-trained network after fine-tuning are 80.05%， which is 8.83 percentage points better than training directly.
