SageMix: Saliency-Guided Mixup for Point Clouds

Data augmentation is key to improving the generalization ability of deeplearning models. Mixup is a simple and widely-used data augmentation techniquethat has proven effective in alleviating the problems of overfitting and datascarcity. Also, recent studies of saliency-aware Mixup in the image domain showthat preserving discriminative parts is beneficial to improving thegeneralization performance. However, these Mixup-based data augmentations areunderexplored in 3D vision, especially in point clouds. In this paper, wepropose SageMix, a saliency-guided Mixup for point clouds to preserve salientlocal structures. Specifically, we extract salient regions from two pointclouds and smoothly combine them into one continuous shape. With a simplesequential sampling by re-weighted saliency scores, SageMix preserves the localstructure of salient regions. Extensive experiments demonstrate that theproposed method consistently outperforms existing Mixup methods in variousbenchmark point cloud datasets. With PointNet++, our method achieves anaccuracy gain of 2.6% and 4.0% over standard training in 3D Warehouse dataset(MN40) and ScanObjectNN, respectively. In addition to generalizationperformance, SageMix improves robustness and uncertainty calibration. Moreover,when adopting our method to various tasks including part segmentation andstandard 2D image classification, our method achieves competitive performance.