A Survey of Multi-Task Learning For Dense Scene Understanding

PhD Qualifying Examination


Title: "A Survey of Multi-Task Learning For Dense Scene Understanding"

by

Mr. Hanrong YE


Abstract:

Multi-Task Dense Scene Understanding is a critical research topic in the 
field of computer vision, which has numerous applications in real-world 
scenarios. To achieve a comprehensive understanding of visual information 
at the pixel level, several fundamental dense prediction tasks, such as 
semantic segmentation, depth estimation, and surface normal estimation, 
are required. Simultaneously learning multiple challenging tasks in a 
unified model is highly demanding, making multi-task dense scene 
understanding a challenging problem. In this survey, we aim to provide a 
systematic review of the developments in this field, particularly from the 
perspective of deep learning. We begin by defining the problem and 
highlighting its importance. We then discuss various existing methods that 
have been proposed to tackle this problem by improving the multi-task 
optimization process and designing multi-task network architecture. We 
examine representative works of each genre from different perspectives. 
Furthermore, we consider a more practical setting: multi-task 
semi-supervised learning, where training samples are labeled for only some 
tasks rather than all tasks. We introduce pioneering works in this 
direction and analyze their advantages and disadvantages. Overall, we hope 
that this survey will serve as a foundation for further research in this 
field and inspire new ideas for tackling this challenging problem.


Date:                   Thursday, 20 April 2023

Time:                   10:00am - 12:00noon

Venue:                  Room 5501
                         Lifts 25/26

Committee Members:      Dr. Dan Xu (Supervisor)
                         Dr. Qifeng Chen (Chairperson)
 			Prof. Huamin Qu
 			Prof. Raymond Wong


**** ALL are Welcome ****