We address the challenge of domain adaptation in LiDAR-based 3D object detection by introducing a simple yet effective training strategy known as Gradual Batch Alternation. This method enables adaptation from a well-labeled source domain to an insufficiently labeled target domain. Initially, training commences with alternating batches of samples from both the source and target domains. As the training progresses, we gradually reduce the number of samples from the source domain. Consequently, the model undergoes a gradual transition towards the target domain, resulting in improved adaptation. Domain adaptation experiments for 3D object detection on four benchmark autonomous driving datasets, namely ONCE, PandaSet, Waymo, and nuScenes, demonstrate significant performance gains over prior works and strong baselines.
Gradual Batch Alternation for Effective Domain Adaptation in LiDAR-Based 3D Object Detection
2024 IEEE Intelligent Vehicles Symposium (IV) ; 2213-2219
2024-06-02
1927650 byte
Conference paper
Electronic Resource
English