OL-Aug: online LiDAR data augmentation for 3D detection
Shida Wei, Jiacheng Liu, Rui Ma · 2024
LiDAR point cloud data plays a vital role in autonomous driving systems by enabling essential 3D perception tasks like 3D object detection and segmentation. However, the scarcity of labeled LiDAR data hampers the development of robust deep learning algorithms for these tasks. Data augmentation has been widely used to increase labeled data in various ways, such as geometric transformation, mixup, and inserting synthetic objects. In this paper, we specifically focus on exploring more effective online LiDAR data augmentation techniques. We propose OL-Aug, which contains two online augmentation modules, namely Swap-GT and GT-Aug++, to enhance the realism and usefulness of augmented data. Unlike previous offline LiDAR data augmentation approaches, our Swap-GT module swaps objects in the current scene with the objects which have closest location and size from an object database in an online manner. In addition, the GTAug++ module not only inserts objects from the database but also removes the occluded background point clouds. To evaluate the effectiveness of our proposed OL-Aug approach, we conduct experiments on the KITTI dataset for 3D object detection. The results demonstrate that OL-Aug outperforms previous state-of-the-art LiDAR data augmentation methods.