Training large-scale video generative adversarial networks for high quality video synthesis

Generative Adversarial Networks;video synthesis;computer vision

Du, Xiaodan ; Lazebnik ; Svetlana

关键词: Generative Adversarial Networks;    video synthesis;    computer vision;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/108048/DU-THESIS-2020.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Video synthesis using deep learning methods is an important yet challenging task for the computer vision community.Generative Adversarial Networks have been proved effective for generating high fidelity photo-realistic images. Recently, many video synthesis models achieve high fidelity and resolution samples by carrying the success of Generative Adversarial Networks to the field of video synthesis. However, it can be challenging to train large-scale Generative Adversarial Networks as they often require enormous computing resources and a long training period. We found it necessary to put together a clear and in-depth guideline for researchers who are interested in training large-scale video Generative Adversarial Networks in the future. In this thesis, we aim to find effective and efficient ways to implement and train large-scale video Generative Adversarial Networks for high quality video generation. We evaluate different implement choices as well as training details and give quantitative analysis.