LiDAR-Transfer: A Flexible Data Augmentation for Point Cloud

LiDAR data augmentation plays a crucial role in enhancing the accuracy of LiDAR-based 3D detection tasks. This paper introduces a novel LiDAR data augmentation method called LiDAR-Transfer, which facilitates the transfer of samples from a source dataset to a target dataset. The proposed approach involves a simple and efficient ground truth sample extraction process to create a comprehensive ground truth database. A space selection strategy is then employed to identify suitable positions for placing the extracted ground truth samples in the target dataset. Finally, a rotation and distribution modification technique is applied to ensure the added samples appear more realistic in the target dataset. In comparison to the existing similar method GT-Sampling, LiDAR-Transfer enables sample addition at any position without considering its original location in the source dataset. The effectiveness of our method is validated on popular LiDAR-based detection frameworks such as SECOND, PointPillar, and CenterPoint, using the Waymo dataset. The best experimental results in CenterPoint demonstrate the improvement in performance, achieving an impressive 70.88% mAP on the Waymo validation dataset, which is 0.67% higher than the original model augmented by GT-Sampling.