Geoinformatics Team Seminar (Talk by Aoran Xiao, Nanyang Technological University).

This is an online seminar. Registration is required.
【Geoinformatics Team】
【Date】2024/April/3(Wed) 14:00-15:00(JST)
*【Speaker】Aoran Xiao, Nanyang Technological University *

Title: Label Efficient Learning of 3D LiDAR Point Cloud Segmentation

Abstract:
LiDAR point cloud segmentation provides fine-grained scene understanding from precise 3D spatial data, thereby enabling accurate perception and interpretation in applications such as autonomous driving, urban planning, and environmental monitoring. While deep learning has achieved notable progress in this area, most 3D segmentation models require large-scale point-wise annotations as training supervision, which remain both laborious and expensive for collection. This challenge restricts the scalability of existing LiDAR point cloud datasets and serves as a bottleneck for the effective deployment of LiDAR technology across a multitude of tasks and applications. Label-efficient learning emerges as a promising solution, enabling robust deep network training while markedly reducing the need for extensive annotations. This talk delves into the significance of label-efficient learning within the realm of 3D LiDAR point cloud segmentation. It will focus on three key aspects: data augmentation, domain transfer learning from synthetic to real point clouds, and domain transfer learning from normal-to-adverse weather conditions. A systematic introduction to our pioneering research efforts in this field will be presented, encompassing datasets, algorithms, and literature surveys.