Abstract
While multi-class 3D detectors are needed in many robotics applications,training them with fully labeled datasets can be expensive in labeling cost. Analternative approach is to have targeted single-class labels on disjoint datasamples. In this paper, we are interested in training a multi-class 3D objectdetection model, while using these single-class labeled data. We begin bydetailing the unique stance of our "Single-Class Supervision" (SCS) settingwith respect to related concepts such as partial supervision and semisupervision. Then, based on the case study of training the multi-class versionof Range Sparse Net (RSN), we adapt a spectrum of algorithms -- from supervisedlearning to pseudo-labeling -- to fully exploit the properties of our SCSsetting, and perform extensive ablation studies to identify the most effectivealgorithm and practice. Empirical experiments on the Waymo Open Dataset showthat proper training under SCS can approach or match full supervision trainingwhile saving labeling costs.