Approaches for Point Cloud Object Detection and Transfer Learning
Presentation Abstract:
Autonomous driving is dependent on the 3D detection of objects like vehicles and pedestrians for navigation. The majority of existing methods rely on LiDAR sensors to provide precise depth data with a point cloud of the scene structure. Several Deep Learning architectures have been proposed to tackle this problem. However, they are usually specific to a dataset, the popular ones being KITTI, Waymo, and nuScenes. These datasets capture a diverse set of driving scenarios in different driving conditions. To improve the performance of an architecture for a generalized input, techniques of learning from multiple datasets can be helpful. Thus, this work aims to experiment with transfer learning techniques applied to the KITTI and nuScenes dataset with popular baseline architectures for 3d Object Detection on LiDAR data.
Advisor: Ziming Zhang
Zoom Link: https://wpi.zoom.us/j/98616398146