Unsupervised Domain Adaptation for 3D Object Detection via Self-Training

3D object detection based on point clouds plays a crucial role in autonomous driving. High quality detection results can provide reliable basis for subsequent stages such as trajectory prediction and path planning. Although many advanced 3D object detectors currently exist, when employing them to another domain, there is often a huge performance drop. In addition, existing domain adaptation methods for 3D object detection only focus on one or two observable factors such as scale mismatch and density variation that lead to domain shift and they do not take some invisible factors (weather, road condition and sensor type, etc.) into account. Therefore, we attempt to propose a self-training pipeline for unsupervised domain adaptation on 3D object detection. Firstly, we pretrain the detectors with a specific data processing paradigm which includes object random scaling, random beam re-sampling and etc. Then, we employ mean-teacher framework which includes cross-domain student model and target-only teacher model. We employ adversarial learning in student model, enforcing the student model to learn domain-invariant features. This process could further eliminate the invisible factors that lead to domain shift. Furthermore, in order to further obtain high-quality pseudo labels, we apply different data augmentation strategy and mutual learning between student model and teacher model. In addition, we adopt domain statistics normalization to ensure a stable training behavior. Extensive experiments under three different adaptation tasks demonstrate the effectiveness of our method.