20210609-15 总第241期 领域自适应方法与进展

报告嘉宾：Dengxin Dai (ETH Zurich)

报告题目：Domain Adaptation for Real-world Domain Changes

报告嘉宾：武阿明 (西安电子科技大学)

报告题目：基于渐进解耦的实例不变域自适应目标检测

Panel嘉宾：

Dengxin Dai (ETH Zurich)，武阿明 (西安电子科技大学)、李文 (电子科技大学)、刘子纬 (Nanyang Technological University)

Panel议题：

1. 领域自适应方法目的是解决什么问题？难点是什么？进展如何？瓶颈是什么？

2. 目前的方法通常需要接触大量的目标域数据，在实际场景中，可能没有这么多目标域数据，这种情况下，如何提升模型的泛化能力？

3. 多目标域自适应研究近期也受到很多关注，该问题有哪些特殊的挑战？还有哪些其他值得关注的领域自适应研究方向？

4. 如何确定哪些领域能够自适应，哪些领域自适应难度大？

5. 领域自适应在哪些实际任务、应用上有迫切需求？目前落地的情况如何？

*欢迎大家在下方留言提出主题相关问题，主持人和panel嘉宾会从中选择若干热度高的问题加入panel议题！

报告嘉宾：Dengxin Dai (ETH Zurich)

报告时间：2021年06月09日 (星期三)晚上20:00 (北京时间)

报告题目：Domain Adaptation for Real-world Domain Changes

报告人简介：

Dengxin Dai is a Senior Scientist and Lecturer working with the Computer Vision Lab at ETH Zurich. He leads the research group TRACE-Zurich working on Autonomous Driving in cooperation with Toyota. In 2016, he obtained his PhD in Computer Vision at ETH Zurich. He is the organizer of the workshop series (CVPR'19-21) "Vision for All Seasons: Bad Weather and Nighttime", the ICCV'19 workshop "Autonomous Driving", and the ICCV'21 workshop "DeepMTL: Multi-Task Learning in Computer Vision". He was a Guest Editor for the IJCV special issue "Vision for All Seasons", an Area Chair for WACV 2020, and an Area Chair for CVPR 2021. His research interests lie in Autonomous Driving, Robust Perception Algorithms, Lifelong Learning, Multi-task Learning, and Multimodal Learning.

个人主页：

http://people.ee.ethz.ch/~daid

报告摘要：

In this talk, I will first give an overview of our work on domain adaptation for real-world domain changes such as from clear weather to adverse weather and from daytime to nighttime. In this part, particular focus will be given to our ACDC dataset. ACDC is a new large-scale driving dataset for training and testing semantic image segmentation algorithms on adverse visual conditions, such as fog, nighttime, rain, and snow. The dataset and associated benchmarks are made publicly available. I will then present our recent work on learning with auxiliary tasks for domain adaptation and knowledge transfer. Specifically, I will present our recent methods for semi-supervised and domain-adaptive semantic image segmentation by using self-supervised depth estimation as an auxiliary task. Our methods have achieved state-of-the-art results for both tasks. The codes are available. Finally, I will conclude the talk with another work on Learning Hyperspectral Image Super-Resolution with RGB Image Super-Resolution as the auxiliary task. The method sets a new state of the art on four public benchmarks and the code is available as well.

报告嘉宾：武阿明 (西安电子科技大学)

报告时间：2021年06月09日 (星期三)晚上20:30 (北京时间)

报告题目：基于渐进解耦的实例不变域自适应目标检测

报告人简介：

武阿明，西安电子科技大学讲师，2021年获得天津大学博士学位，2019年至2020年在澳大利亚悉尼科技大学交流学习。主要研究领域为多模态表示与学习，专注于视频描述生成、视觉常识推理、领域适应、小样本学习等关键问题。发表多篇会议、期刊论文，包括TPAMI、TCSVT、NeurIPS、IJCAI等。

个人主页：

https://amingwu.github.io/

报告摘要：

基于深度学习的目标检测已经取得许多进展，然而，当训练数据和测试数据之间存在域偏移时，算法的有效性和稳定性会面临巨大挑战。对于这个问题，提取域不变的实例层面特征（即实例不变特征）是一种有效的解决方法。因此，我们探索通过将域不变和域特定特征进行分离来提取实例不变特征。具体地，我们提出了一个渐进解耦的框架。它主要由基础解耦层和渐进解耦层组成。其中，通过将域不变特征与当前特征进行融合，基础解耦层能够增强网络低层特征的域不变信息，从而有助于准确地提取候选物体。接下来，渐进解耦层旨在基于分解的域不变特征来提取实例不变特征。在实例不变特征空间中进行定位和分类可以消除域偏移对模型的影响。多种域偏移场景的实验结果表明了所提方法的有效性。

Panel嘉宾：李文 (电子科技大学)

嘉宾简介：

李文，电子科技大学教授，博士生导师，四川省千人特聘专家，2015年获新加坡南洋理工大学博士，2015年至2019年在瑞士苏黎世联邦理工学院计算机视觉实验室从事研究工作。主要研究领域为计算机视觉与机器学习，专注于计算机视觉任务中的领域适应、迁移学习、弱监督学习、半监督学习等关键问题，在T-PAMI、IJCV、CVPR、ICCV、ECCV等在内的领域重要国际期刊和国际会议论文40余篇，Google Scholar的总引用次数3600余次。是迁移学习领域重要研讨会ICCV/ECCV Workshops on TASK-CV的主办者之一，以及互联网数据学习研讨会CVPR Workshops on WebVision的发起人之一。

长期担任包括T-PAMI、T-IP、T-NNLS、CVPR、ICCV、ECCV、NeurIPS、ICML、ICLR在内的重要国际期刊和会议审稿人，担任AAAI 2021领域主席、ACM MM 2021注册主席，获ECCV 2016、CVPR 2019杰出审稿人奖，与苏黎世联邦理工学院、南洋理工大学、悉尼大学、谷歌等一流科研机构和企业长期保持紧密的合作关系。

个人主页：

https://wenli-vision.github.io/

Panel嘉宾：刘子纬 (Nanyang Technological University)

嘉宾简介：

Prof. Ziwei Liu is currently a Nanyang Assistant Professor at Nanyang Technological University (NTU). Previously, he was a senior research fellow at the Chinese University of Hong Kong. Before that, Ziwei was a postdoctoral researcher at University of California, Berkeley, working with Prof. Stella Yu. Ziwei received his PhD from the Chinese University of Hong Kong in 2017, under the supervision of Prof. Xiaoou Tang and Prof. Xiaogang Wang. During his PhD, Ziwei had the privilege of interning at Microsoft Research and Google Research, where he developed Microsoft Pix and Google Clips. His research revolves around computer vision/graphics, machine learning, and robotics. He has published over 70 papers (with more than 9,700 citations) on top-tier conferences and journals in relevant fields, including CVPR, ICCV, ECCV, ICLR, IROS, SIGGRAPH, TPAMI, TOG, and Nature-Scientific Reports. He serves as an Area Chair of ICCV 2021 and the Associate Editor of IET Computer Vision. He is the recipient of Microsoft Young Fellowship, Hong Kong PhD Fellowship, ICCV Young Researcher Award, and HKSTP best paper award. He has won the championship in major computer vision competitions, including DAVIS video segmentation challenge 2017, MSCOCO instance segmentation challenge 2018, FAIR self-supervision challenge 2019, and Video Virtual Try-on Challenge 2020. He is also the lead contributor of several renowned computer vision benchmarks and softwares, including CelebA, DeepFashion, mmdetection and mmfashion.

个人主页：

https://liuziwei7.github.io

主持人：朱霖潮 (悉尼科技大学)

主持人简介：

朱霖潮，悉尼科技大学讲师。分别于浙江大学和悉尼科技大学获得本科与博士学位，2015年和2016年于卡内基梅隆大学访学。曾获得美国国家标准总局TRECVID比赛冠军，EPIC-Kitchens，THUMOS动作识别比赛冠军。2021年获Google Research Scholar奖（在机器感知领域仅有七个获奖者）。长期关注视频行为理解，无监督视频特征学习，元学习等。

个人主页：

ffmpbgrnn.github.io/

21-15期VALSE在线学术报告参与方式：

长按或扫描下方二维码，关注“VALSE”微信公众号 (valse_wechat)，后台回复“15期”，获取直播地址。

特别鸣谢本次Webinar主要组织者：

主办AC：朱霖潮 (悉尼科技大学)

协办AC：李文 (电子科技大学)