LiDAR-camera-system-based unsupervised and weakly supervised 3D object detection

Abstract

LiDAR camera systems are now becoming an important part of autonomous driving 3D object detection. Due to limitations in time and resources, only a few critical frames of the synchronized camera data and acquired LiDAR points may be annotated. However, there is still a large amount of unannotated data in practical applications. Therefore, we propose a LiDAR-camera-system-based unsupervised and weakly supervised (LCUW) network as a novel 3D object-detection method. When unannotated data are put into the network, we propose an independent learning mode, which is an unsupervised data preprocessing module. Meanwhile, for detection tasks with high accuracy requirements, we propose an Accompany Construction mode, which is a weakly supervised data preprocessing module that requires only a small amount of annotated data. Then, we generate high-quality training data from the remaining unlabeled data. We also propose a full aggregation bridge block in the feature-extraction part, which uses a stepwise fusion and deepening representation strategy to improve the accuracy. Our comparative, ablation, and runtime test experiments show that the proposed method performs well while advancing the application of LiDAR camera systems.