Mix3D data augmentation enhanced RandLA-Net for large-scale point cloud semantic segmentation

3D point cloud semantic segmentation is one of the key technologies for fast 3D modeling of digital twins. To address the overfitting problem of the large-scale semantic segmentation model RandLA-Net, this paper proposes an improved RandLA-Net method based on Mix3D data augmentation. RandLA-Net is a high-volume, large-perception field model that can directly capture the contextual information of the entire 3D scene, while 3D datasets tend to be more expensive due to data sampling and labeling, often the number of scenes is small and the variance within the data is small, RandLA-Net can easily learn the overly strong contextual prior on the training set, and the model may show poor generalization ability when reasoning in realistic scenes. By introducing Mix3D to mix the two scenes to generate new training samples and implicitly place the object instances in the new contextual environment, the RandLA-Net model no longer relies solely on the scene context to infer semantic labels, but instead infer from the local structure, balancing the role of global context and local structure information in model inference and effectively reducing the overfitting of the training set context. The overfitting of the training set context is effectively reduced. Experimental results on several datasets show that our approach results in a 1.3% and 0.6% mIoU improvement of the RandLA-Net model on Semantic3D and S3DIS datasets.