A dual projection method for semantic segmentation of large-scale point clouds

3D point cloud semantic segmentation is an essential task for enhancing high-level perception in autonomous platforms. The projection-based method is an efficient approach since central projection can transform sparse point clouds into dense pseudo-2D image. Central projection suffers from scale imbalances, posing challenges for perceiving distant objects. We utilized parallel projection to adjust the distribution and scale of targets, enabling neural networks to better extract information from distant objects. We designed a joint training and inference framework to combine the perspectives of central projection and parallel projection, enhancing the final segmentation results. Our method achieves an mAcc of 89.7% and an mIoU of 59.3% on the SemanticKITTI validation set. On the SemanticKITTI test set, we achieve an mAcc of 89.0% and an mIoU of 58.7%. The experiments show that our method achieves superior segmentation performance, especially for distant targets and categories such as motor vehicles, distant road surfaces and column-like structures.