PointCutMix: Regularization strategy for point cloud classification

As 3D point cloud analysis has received increasing attention, the insufficient scale of datasets and weak generalization ability networks become prominent. In this paper, we propose a simple effective augmentation method for data to alleviate those problems. It finds one-to-one correspondence between two clouds generates new training by replacing points in one sample with their corresponding pairs. Two replacement strategies are proposed adapt accuracy or robustness requirement different tasks, which is randomly select all while other k nearest neighbors single random point. Both consistently significantly improve performance various models on classification By introducing saliency maps guide selection points, further improves. Since our persists local semantic information also at first time extend MSDA segmentation problem. Moreover, PointCutMix validated enhance model’s robustness. When using as defense method, outperforms state-of-the-art algorithms. The code available at:https://github.com/cuge1995/PointCutMix.