AdaMix: Adaptive Resampling of Multiscale Object Mixup for Lidar Data Augmentation

Lidar data, which can describe the 3D spatial information of the environment in the form of point clouds, play an important role in autonomous driving and related downstream tasks such as 3D object detection. However, unlike for images, collecting and labeling lidar data is often very expensive. As an effective means to increase the quantity of annotated data for training deep learning models, data augmentation (DA) has been widely used in the image field, but studies on augmenting lidar point clouds are only at the beginning stage. In this article, we propose AdaMix, a novel framework for lidar DA via adaptive resampling of multiscale object mixup. AdaMix contains two different object mixup schemes, i.e., object-level and part-level mixup, to augment the lidar data with the existing object instances from different scenes. For object-level mixup, a learning-based point upsampling operation is employed to obtain a set of dense objects, such as vehicles and pedestrians. For part-level mixup, parts from different vehicles are composed together and upsampled to generate vehicles of complete and dense shapes. To mix the dense objects into a new scene, AdaMix introduces a novel projection-based downsampling method to adaptively downsample the objects based on the location generated from a location sampling module. We evaluate the performance of AdaMix with several 3D object detection models on the KITTI dataset. Experimental results demonstrate that AdaMix consistently surpasses state-of-the-art lidar DA methods in improving the average precision of vehicle and pedestrian detection.