LEARNING TO ENFORCE AND UTILIZE TEMPORAL CONSISTENCY IN VIDEO PROCESSING

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "LEARNING TO ENFORCE AND UTILIZE TEMPORAL CONSISTENCY IN VIDEO 
PROCESSING"

By

Mr. Chenyang LEI


Abstract

Countless videos exist in the world with diverse contents and styles. 
However, in many cases, the original videos captured or created are not 
perfect, and many video processing algorithms are proposed to process 
these diverse videos for various purposes. Video temporal consistency is a 
common goal for various video processing algorithms, while these 
algorithms are designed for different downstream applications. The video 
temporal consistency denotes the property that correspondences of 
consecutive frames in a video share the consistent features (e.g., color). 
While this property exists in most natural videos, it might be destroyed 
when videos are processed by algorithms. Video temporal consistency in 
video processing is challenging since it is related to several hard 
problems in computer vision, including correspondences estimation, 
task-specific reconstruction, and so on. This thesis focuses on studying 
the video temporal consistency problem with machine learning. Since 
temporal consistency is a common goal for video processing algorithms, can 
deep networks learn the temporal consistency from large-scale data or few 
data? We try to propose several approaches that can obtain satisfying 
performance on various video processing tasks.


Date:			Friday, 26 August 2022

Time:			10:00am - 12:00noon

Zoom Meeting:
https://hkust.zoom.us/j/96298250397?pwd=YXV6cTdCYkd5djk4ZkRsYkcwRm94Zz09

Chairperson:		Prof. Daniel PALOMAR (ECE)

Committee Members:	Prof. Qifeng CHEN (Supervisor)
 			Prof. Pedro SANDER
 			Prof. Dan XU
 			Prof. Ling SHI (ECE)
 			Prof. Hongsheng LI (CUHK)


**** ALL are Welcome ****