Optimising LiDAR-Based 3D Object Detection with Limited Labels

The School of ITEE is hosting the following PhD confirmation progress seminar:

Optimising LiDAR-Based 3D Object Detection with Limited Labels

Speaker: Zhuoxiao (Ivan) Chen
Host: Dr Yadan Luo

Abstract: LiDAR-based 3D object detection plays an indispensable role in 3D scene understanding with a wide range of applications such as autonomous driving and robotics. Over the past few years, there has been a rise in the number of approaches for 3D object detection, achieving outstanding performance. However, these methods depend considerably on being trained on large-scale 3D datasets with numerous manually labelled annotations. In order to overcome this difficulty, our research focuses on optimising the model performance for 3D object detection, especially when the labelling budget is limited. Our first work derives a novel active learning (AL) technique that learns to select only a small portion of unlabelled data of which the label is concise, feature is representative and geometrical characteristic. Annotations obtained for the selected subset contribute to an improvement in detection performance. To improve the generalizability of the 3D detectors for out-of-domain scenes without accessing labels, we explore domain-adaptive approaches by generating reliable, diverse and class-balanced pseudo labels from unlabelled and unseen data. Extensive experiments evidence the effectiveness of our methods for comprehending both in- and out-of-domain 3D scenes, requiring only the least amount of human annotation.

Speaker: Zhuoxiao (Ivan) Chen is a PhD student from the Data Science group at the School of Information Technology and Electrical Engineering, the University of Queensland (UQ), Australia. He received his Bachelor of Computer Science (Honours) degree from UQ. He is currently working towards his PhD degree under the supervision of Dr Yadan Luo, Prof Helen Huang,  and Dr Mahsa Baktashmotlagh. His research interests include active learning, domain adaptation for 3D scene understanding.