Unsupervised Objects Discovery from LiDAR Scans for Autonomous Driving

Abstract

Self-driving promises to revolutionize the current transportation paradigm, with the potential of providing a safer and more efficient solution for transportation. A critical component for safe self-driving lies in its ability to perceive the world and sense its surroundings. Despite the rapid progress of 3D perception algorithms designed for this field, a majority rely on supervised training from 3D bounding box labels which are both expensive and labor intensive to obtain. The aim of my research is to shift the current paradigm away from these supervised training methods, and instead train 3D object detectors without any labels at all. In this talk, I will briefly go over the field of label-free object detection. I will then present studies into how this can be achieved, and discuss future research directions that I will pursue in this area.