Biography: Dr. Xun Xu received his PhD from Queen Mary University of London in 2016. He then worked as research fellow with National University of Singapore from 2016-19. He joined A*STAR as research scientist from 2019. Dr. Xu is interested in developing data-efficient and robust learning methods for computer vision problems. He has published over 30 papers, including IEEE TPAMI, IJCV, CVPR, NeurIPS, ECCV, IEEE TIP, etc. He has received one AME Programmatic Fund as Co-I (SGD$1.4M), one Career Develop Award (SGD$23.8k) and one NSFC Youth grant (RMB$300k).

Title: Data-Efficient 3D Point Cloud Deep Learning

Abstract: Humans live in a 3D space, thus understanding the world in 3D is the fundamental task for computer vision research. Point cloud is an efficient 3D representation and can be acquired through a wide range of sensors and algorithms, e.g. LiDAR, structured light, structure from motion. Understanding 3D point cloud requires recognition, segmentation and detecting objects of interest from unstructured 3D environments and this is often achieved by training deep neural networks (DNN). To obtain DNN with higher performance it often requires large labelled data, which is expensive to acquire. Thus, it has motivated research into data-efficient learning, including learning with fewer labelled data, exploiting unlabelled data, selecting most informative data, etc. This talk will introduce research works targeting data-efficient 3D point cloud deep learning from multiple perspectives. The developed techniques have benefited 3D point cloud recognition, semantic segmentation, and object detection tasks. It has been also demonstrated on autonomous driving applications by providing more efficient data annotation strategies.