Title: Representation learning for object detection from unlabeled LiDAR point cloud sequences
Abstract: There are a huge number of LiDAR point cloud sequences that are generated everyday and a large fraction of them never get annotated. We will discuss how to use unlabeled LiDAR point cloud sequences in a way that requires no box annotations. The key observation is to look for "objects that are moving along smooth trajectories", or object traces.
Such intermediate data can be reliably extracted from LiDAR point cloud sequences without any learning techniques and is valuable by itself. For the downstream task, we design self-supervised pretext tasks that improve the performance of object detection.
Bio: Xiangru Huang received his PhD in Computer Science as a student of Qixing Huang from the University of Texas at Austin in 2020. His past research focused on efficient optimization algorithms, geometry processing and machine learning. In 2021, he then joined the Geometric Data Processing group as a postdoc, working with Prof. Justin Solomon. His more recent research focuses on self-supervised learning and point cloud sequences.