Class-Balanced PolarMix for Data Augmentation of 3D LIDAR Point Clouds Semantic Segmentation

Abstract

3D LIDAR point clouds are extensively utilized in various domains, and data augmentation techniques for these point clouds can enhance network model convergence during training while also reducing the requisite data volume. Notably, PolarMix represents a seminal contribution to data enhancement in the realm of 3D LIDAR point Clouds Semantic Segmentation. It markedly augments the number of instances per class through swapping and rotate-paste mechanisms. Rotate-paste involves rotating and pasting selected class instances around the Z-axis multiple times. However, when capturing real-world scenarios using LiDAR point clouds, a pronounced class imbalance is observed, wherein certain classes dominate in sample numbers, while others are sparsely represented. Regrettably, PolarMix overlooks this class imbalance, leading to unequal treatment of all classes. To rectify this, we introduce the Class-Balanced PolarMix (CB-PolarMix), which operates in a cascading manner to diversify the training distribution and further optimize data augmentation outcomes. The cornerstone of CB-PolarMix lies in its adaptive reinforcement of foreground classes based on their distribution patterns. More specifically, our approach tweaks the pasting process for each class contingent upon its historical prediction accuracy. Experimental results from the SemanticPOSS and SemanticKitti datasets, utilizing the MinkowskiNet and SPVCNN models respectively, underscore the efficacy of the proposed CB-PolarMix.