报告题目：Generative Adversarial Networks for Cross-domain Visual Data
报告人： Chang Xu
Dr Chang Xu is a Lecturer in Machine Learning and Computer Vision at the School of Computer Science, University of Sydney. He obtained a Bachelor of Engineering from Tianjin University, China, and a Ph.D. degree from Peking University, China. While pursuing his PhD degree, Chang received fellowships from IBM and Baidu. His research interests lie in machine learning, data mining algorithms and related applications in artificial intelligence and computer vision, including multi-view learning, multi-label learning, visual search and face recognition. His research outcomes have been widely published in prestigious journals and top conference.
This talk will introduce generative adversarial networks (GAN) for cross-domain visual data. In particular, we focus on two kinds of tasks, image-to-image translation and image-to-video translation. In the first task, to relieve the burden of classical GAN, we develop a new attention GAN technique, which decomposes the generative network of classical GAN into two separated networks: the attention network predicting spatial attention maps of images, and the transformation network focusing on translating objects. To address the second task, we propose a two-stage unsupervised motion adaptation framework. We first extract motion information from source video clips, and then the learned motions are combined with different target image to synthesize desired videos. We devised a motion adaptation generative adversarial network and an image appearance adversarial network to evaluate the correctness of the synthesized frame’s motion and appearance, respectively. Experimental results on real-world datasets demonstrate the effectiveness of the proposed algorithms.