VALSE Webinar 19-07期 迁移学习与领域适配

报告人简介：

李文，目前任瑞士苏黎世联邦理工学院计算机视觉实验室博士后研究员，合作导师为知名计算机专家Luc Van Gool教授。他于2015年在新加坡南洋理工大学取得博士学位，主要研究方向为计算机视觉与迁移学习，重点研究视觉应用中的标注数据有限、数据分布差异等问题，在T-PAMI、IJCV、CVPR、ICCV、ECCV等重要国际期刊和会议上发表30多篇学术论文，谷歌学术引用1400余次。

个人主页：

http://www.vision.ee.ethz.ch/~liwenw/

报告摘要：

Deep learning models often suffer from the challenge of domain shift in real world scenarios. That is, the test data (i.e., the target domain) exhibits a considerable distribution difference from the training data (i.e., the source domain). In this talk, I will present our recent two works on using image translation techniques to generate training data for improving the model performance in the target domain. In the first work, we propose a new domain flow generation (DLOW) approach, which is able to translate images from the source domain into an arbitrary intermediate domain between source and target domains. The generated data in the intermediate domains alleviate the domain difference between source and target domains, thus can be used to improve the cross-domain generalization ability of deep learning model. In the second work, we improve the conventional image translation method by considering privileged information in source domain. Taking synthetic to real domain adaptation as an example, we show that the geometric information from synthetic data helps to generated high-quality real style images, and therefore improves the model performance when those generated real style images are used as the training data.

参考文献：

[1] Yuhua Chen, Wen Li, Xiaoran Chen, Luc Van Gool. Learning Semantic Segmentation from Synthetic Data: A Geometrically Guided Input-Output Adaptation Approach. CVPR 2019.

[2] Rui Gong,  Wen Li,  Yuhua Chen,   Luc Van Gool. DLOW: Domain Flow for Adaptation and Generalization. CVPR 2019.

报告嘉宾： 邓成（西安电子科技大学）

报告时间： 2019年4月3日（星期三）晚上20 :30 （北京时间）

报告题目： Zero-shot Learning for Cross-domain Information Retrieval