Evaluation of Point Cloud Data Augmentation for 3D-LiDAR Object Detection in Autonomous Driving

Obstacle detection is an essential component for autonomous vehicles to navigate safely. To address certain limitations of 2D object detection, numerous recent approaches have emerged, aiming to leverage LiDAR sensors for 3D object detection. LiDAR sensors, with their depth sensing capabilities, offer richer spatial and geometric data, enabling the accurate estimation of 3D bounding boxes and object orientations. However, the complexity of 3D annotation, in supervised object detection, imposes significant challenges for this task. To address the annotation problem, some studies have explored data augmentation techniques for 3D point clouds, but, not all augmentation methods yield positive impacts on model performance. Therefore, this paper presents an in-depth evaluation of global data augmentation techniques, specifically focusing on geometric transformation and noise-based methods for 3D object detection. The results reported in this paper, achieved on the KITTI dataset, showed a relevant difference in some geometric operations, and the importance of noise-based methods.