Weakly Supervised Semantic Segmentation: A Survey

PhD Qualifying Examination


Title: "Weakly Supervised Semantic Segmentation: A Survey"

by

Mr. Zhihan GAO


Abstract:

Many applications related to computer vision require efficient understanding of 
input images and videos. Semantic segmentation is to understand images and 
videos at pixel level, i.e., to make dense predictions of the class labels for 
each pixel. Over the past few years, deep learning methods have achieved a 
great success in this field and have made it a much more popular research 
topic. However, manually annotated pixel-level masks, which are extremely 
expensive and require experts, are highly demanded for model training. Thus, it 
is important and promising to explore the training strategies of semantic 
segmentation models with various forms of weak supervision, including image 
tags, bounding boxes, scribbles, etc. Currently most of the weakly supervised 
semantic segmentation methods are based on the network architectures that 
proposed for fully supervised learning, while differ a lot in training 
segmentation networks using weak labels. In this survey, we first provide a 
review about semantic segmentation, including the existing state-of-the-art 
methods of image semantic segmentation. Then we review the existing weakly 
supervised semantic segmentation methods and categorize them according to the 
types of weak supervision. Finally, we make a summary and point out some 
potential research directions.


Date:			Monday, 21 May 2018

Time:                  	10:00am - 12:00noon

Venue:                  Room 5508
                         Lifts 25/26

Committee Members:	Prof. Dit-Yan Yeung (Supervisor)
 			Prof. Chi-Keung Tang (Chairperson)
 			Prof. Albert Chung
 			Dr. Yangqiu Song


**** ALL are Welcome ****