Perceiving geometric structures of objects in images for autonomous vehicles

PhD Qualifying Examination


Title: "Perceiving geometric structures of objects in images for autonomous 
vehicles"

by

Mr. Zhenhua XU


Abstract:

Perception is one of the fundamental tasks of autonomous driving. The 
autonomous vehicle should be able to detect objects in the scene for later 
navigation tasks. Most current works perceive objects in instance-level or 
pixel-level from images, but cannot efficiently obtain their geometric 
structures, which are critical to perception tasks of autonomous vehicles. For 
example, the topology of line-shaped objects like lanes and curbs is important 
to define the drivable area on the road. Usually, the geometric structures of 
objects are represented by graphs, which can either be skeletons or contours. 
Past works detecting the geometric structures of objects by conventional 
methods tend to be lack of generalization and robustness. Therefore, perceiving 
objects with their geometric structures from images by deep learning based 
methods becomes a frontier problem that is worthy to be explored. Such a 
problem can be formulated as an image-to-graph learning problem. In this paper, 
we survey past works related to this problem. First, we define several key 
concepts and formulate the problem in a formal way. Then, we survey past works 
that aim to perceive geometric structures of objects, including conventional 
methods and deep learning based methods. And finally, we list some potential 
research directions in the future.


Date:			Monday, 8 June 2020

Time:                  	4:00pm - 6:00pm

Zoom meeting:           https://hkust.zoom.us/j/95520274146

Committee Members:	Prof. Huamin Qu (Supervisor)
  			Dr. Ming Liu (Supervisor)
 			Prof. Dit-Yan Yeung (Chairperson)
 			Dr. Qifeng Chen


**** ALL are Welcome ****