Elastic Deformable Augmentation for Autonomous Driving

Jungwook Shin, Jaeill Kim, Wonjong Rhee

Abstract: Unlocking the information concealed in 3D point clouds by LiDAR is a key mission for autonomous driving. However, this task is challenging due to the sparsity, continuity, and unordered nature of point clouds. In addition, creating annotated data is time-consuming and expensive. This emphasizes the crucial necessity of data augmentation. Conventional object data augmentation methods such as rotation, scaling, and translation are not fully effective for 3D data. Therefore, we propose a novel augmentation method, Elastic Deformable Augmentation (EDA), which enhances data diversity for better model robustness. EDA applies deformation methods from 3D graphics to the objects, diversifying their shapes without violating occlusion or intensity properties. We demonstrate EDA on the KITTI dataset, where it improves object detection performance, particularly increasing the mean AP scores by 1.55% and 1.00%, respectively. Consequently, our research provides compelling evidence that EDA is a promising approach for augmenting 3D object detection tasks in autonomous driving.