DEEP LEARNING IN IMAGE MATCHING: A SURVEY

PhD Qualifying Examination


Title: "DEEP LEARNING IN IMAGE MATCHING: A SURVEY"

by

Mr. Hongkai CHEN


Abstract:

Image matching aims to establishing reliable correspondences across images 
under perspective and illumination changes, which lays the foundations for 
a wide range of downstream applications, including Structure-from-Motion 
(SfM), Simultaneous Localization and Mapping (SLAM) and Multiview 
Stereo(MVS). Despite the long-standing dominance of conventional image 
matching pipelines, which comprises of hand-crafted local descriptors, 
heuristic pruning strategy and robust geometry estimator, it has been 
shown by recent study, that the matching quality could be remarkably 
boosted by introducing modern deep learning techniques.

In this survey, we will introduce recent progress in learning-based image 
matching. More specifically, we will cover works in two aspects, 1) 
learnable keypoint and local descriptor, 2) learnable matching strategy 
and, where performance of these works will be evaluated from different 
levels. We also introduce efficient and effective methods we have 
proposed, which achieves competitive performance against state-of-the-arts 
in a much lower cost.

With intensive reviews on previous works and experiments, we propose both 
insights and unsolved problems, which we hope to inspire future works.


Date:			Tuesday, 8 June 2021

Time:                  	4:00pm - 6:00pm

Zoom meeting: 		https://hkust.zoom.us/j/7668256778

Committee Members:	Prof. Long Quan (Supervisor)
 			Prof. Chiew-Lan Tai (Chairperson)
 			Dr. Qifeng Chen
 			Dr. Xiaojuan Ma


**** ALL are Welcome ****