20180822-26 赵健：基于深度学习的人物图像理解：人脸识别与人物解析 ... ... ...

报告嘉宾：赵健（新加坡国立大学）

报告时间：2018年08月22日（星期三）晚上20:00（北京时间）

报告题目：基于深度学习的人物图像理解：人脸识别与人物解析（Deep Learning for Human-Centric Image Understanding: Face Recognition & Human Parsing）

主持人：郭裕兰（国防科大）

报告人简介：

赵健，2012年获得北京航空航天大学自动化专业学士学位（导师：董邵鹏，袁梅），2014年获得国防科技大学计算机科学与技术专业硕士学位（导师：陈旭灿），目前受中国留学基金委资助在新加坡国立大学工程学院电子与计算机工程系LV组攻读博士学位（导师：冯佳时，颜水成）。迄今已发表学术论文20余篇，包括以下顶级期刊/会议：T-PAMI，T-IP，NIPS，CVPR，IJCAI，ECCV，ACM MM，BMVC。曾获得ICCV 2017 MS-Celeb-1M大规模人脸识别Hard Set竞赛、Random Set竞赛与Low-Shot Learning竞赛三项冠军，CVPR 2017 LIP人物解析竞赛与人物姿态估计竞赛两项亚军，美国国家标准技术研究所（NIST）2017 IJB-A人脸确认与人脸鉴别两项竞赛冠军。曾担任T-MM，T-IFS，NIPS，ACM MM，AAAI，ICLR等本领域主流期刊/会议的审稿人。曾组织CVPR 2018群体场景下人物图像理解Workshop以及相应的细粒度多人解析与姿态估计竞赛。曾担任ECCV 2018计算机视觉紧致、高效特征表示学习Workshop程序委员会成员。其研究领域涉及深度学习、模式识别、计算机视觉与多媒体分析，研究课题主要专注于基于深度学习的人物图像理解的模型与算法研发，应用于人脸识别、图像生成与人物解析。

Mr. Jian Zhao received his Bachelor’s degree from School of Automation Science and Electrical Engineering, Beihang University, China in 2012 (supervisors: Dr. Shaopeng Dong, Dr. Mei Yuan), and his Master’s degree from School of Computer, National University of Defense Technology, China in 2014 (supervisor: Dr. Xucan Chen). He is currently supported by China Scholarship Council (CSC) to pursue his Ph.D. degree with the Learning and Vision Group, Department of Electrical and Computer Engineering, Faculty of Engineering, National University of Singapore (supervisors: Dr. Jiashi Feng, Dr. Shuicheng Yan). He has published several cutting-edge papers on unconstrained/large-scale/low-shot face verification/identification and human parsing as the first author (including the following CCF-A conferences and journal: NIPS, CVPR, IJCAI, ACM MM; T-PAMI, CCF-B conference: ECCV, and CCF-C conference: BMVC). He has won the top-3 awards several times on world-wide competitions on face recognition, human parsing and pose estimation as the first author (No.1 on ICCV 2017 MS-Celeb-1M Large-Scale Face Recognition Hard Set Challenge, Random Set Challenge and Low-Shot Learning Challenge; No.2 on CVPR 2017 Visual Understanding of Humans in Crowd Scene & the 1st LIP Challenge Human Parsing task and Human Pose Estimation task; No.1 on NIST 2017 IJB-A Face Verification Challenge and Face Identification Challenge). He has served as reviewer for the CCF-A conferences of NIPS, ACM MM, AAAI, and CCF-A journals of T-MM, T-IFS. He has served as the organizer of the CVPR 2018 Workshop on Visual Understanding of Humans in Crowd Scene (VUHCS), and MHP Challenges on Multi-Human Pose Estimation and Fine-Grained Multi-Human Parsing. He has served as the program committee member of the ECCV 2018 Workshop on Compact and Efficient Feature Representation Learning in Computer Vision (CEFRL). His main research interests include deep learning, pattern recognition, computer vision and multimedia. In particular, his research is focused on developing deep neural network models and algorithms for human-centric image understanding, applied to face recognition, image generation and human parsing.

报告摘要：

当前，人们正生活在数据爆炸式增长的时代。随着社会与科技的快速发展进步，人们可以通过越来越多的智能化设备获取到想要的数据。在缤纷复杂的大数据中，视频与图像可能是最贴近人们日常生活的数据形式之一。因此，研发用于人物图像理解的人工智能技术对于便利人们的生产、生活具有重大意义。举一个例子，如果超市的收银员此刻正神情紧张地把他的双手举过头顶，那么很可能有抢劫等异常情况发生，安保希望能够通过实时监控来识别出可疑人物的身份并获取更多信息，那么该如何解决这个问题？另一个例子，一个小女孩看到一张照片里的女明星穿的裙子很漂亮，也想买一件同款，但她不知道裙子的品牌，只能通过这张图片在购物网站以图搜图，怎样能够让购物网站有效检测并识别出图片中的这件裙子并且给女孩推荐同款或一些类似的款式？近年来，深度学习算法和技术已经在学术界与工业界的众多领域取得了诸多突破性进展。在计算机视觉领域，深度学习算法和技术在很多基准数据集都极大改善并提升了人物图像理解的性能。然而，在涉及到视频监控、安防、电子商务、群体行为分析以及自动驾驶等现实场景下，人物图像理解的性能还是不尽如人意。有关问题还需不断付出努力、投入研究，不断做出改进与完善，寻求更优解决方案。讲者博士期间的课题主要专注于基于深度学习的人物图像理解研究，主要分为两个层面展开：人脸识别，用于探索人物的身份信息；人物解析，用于探索人物的细粒度语义信息。本次报告将介绍讲者在人脸识别与人物解析方面的几个有代表性的工作，主要包括：DA-GAN，发表于IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI，影响因子：9.455) 2018，获得2017年美国国家标准技术研究所（NIST）组织的IJB-A人脸确认与人脸鉴别两项竞赛冠军；PIM和3D-PIM，分别发表于CVPR 2018和IJCAI 2018，在Multi-PIE, CFP, IJB-A, LFW等数据集取得当前已发表工作中的最优姿态鲁棒人脸识别性能；ICCV 2017 MS-Celeb-1M大规模人脸识别Hard Set竞赛、Random Set竞赛与Low-Shot Learning竞赛三项冠军的获奖工作；CVPR 2017 LIP人物解析竞赛的获奖工作；MHP v1.0与MHP v2.0，讲者及联合作者首次提出多人解析任务、相应的数据集、评测标准与参考方法，并组织了CVPR 2018群体场景下人物图像理解Workshop以及相应的细粒度多人解析与姿态估计竞赛，为业界相关技术的发展与进步做出了突出贡献。

Currently, we are living in a data explosion era. With the fast development of our society and technology, we are having more and more intelligent devices that are able to capture the data for specific purposes. Among such big data, images and videos might be the most common types that are highly related to our daily life. Thus, developing techniques with artificial intelligence for human-centric image understanding is of great significance for bringing convenience to people’s life and production. For example, if the cashier is lifting his hands nervously, something must be happened in the supermarket. The security might want to check the identity of the man in front of the cashier through this image for more relevant details. How to do this? Another example, if a female celebrity is wearing a very beautiful dress, and a girl wants to buy a same one for herself. She doesn’t know the brand and she might search it on a shopping website with this image. How will the shopping website recognize this dress and recommend the same or similar ones for the girl? Recently, deep learning techniques have made great breakthroughs in many area both academically and industrially. In computer vision, advances of deep learning approaches have remarkably boosted the performance of human-centric image understanding. Several approaches claim to have achieved or even surpassed human performance on several benchmarks. However, in the real-world scenario involving video surveillance, security, e-commerce, group behavior analysis and autonomous driving, the performance for human-centric image understanding is still far from being satisfactory. It is still an open problem in the deep learning community, with many researchers all over the world devoting their efforts on the possible solutions. In my Ph.D. research, I consider the problem of deep learning for human-centric image analysis, and I address this problem from two main perspectives: face recognition for exploring the identity information and human parsing for exploring the fine-grained semantic body information. In this talk, I will introduce some of my representative works on face recognition and human parsing, including DA-GAN, which was published in T-PAMI 2018 and was the foundation of our winning entry to NIST IJB-A Unconstrained Face Recognition Challenge on Verification and Identification; PIM and 3D-PIM, which were published in CVPR 2018 and IJCAI 2018, respectively, achieved the state-of-the-art performance on Multi-PIE, CFP, IJB-A and LFW benchmarks for pose-invariant face recognition; the winning works on ICCV 2017 MS-Celeb-1M Large-Scale Face Recognition Hard Set Challenge, Random Set Challenge and Low-Shot Learning Challenge; the winning work on CVPR 2017 LIP Challenge Human Parsing task; the works of MHP v1.0 and MHP v2.0, by which we proposed a new Multi-Human Parsing task, corresponding datasets, evaluation metrics and baseline methods, and organized the CVPR 2018 Workshop on Visual Understanding of Humans in Crowd Scene (VUHCS 2018) and the MHP Challenges on Multi-Human Pose Estimation and Fine-Grained Multi-Human Parsing.

参考文献：

[1] 3D-Aided Dual-Agent GANs for Unconstrained Face Recognition, Jian Zhao, Lin Xiong, Jianshu Li, Junliang Xing, Shuicheng Yan, Jiashi Feng, IEEE Transactions on Pattern Analysis and Machine Intelligence (T-PAMI), 2018.

[2] Towards Pose Invariant Face Recognition in the Wild, Jian Zhao, Yu Cheng, Yan Xu, Lin Xiong, Jianshu Li, Fang Zhao, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, Shuicheng Yan, Jiashi Feng, IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018.

[3] 3D-Aided Deep Pose-Invariant Face Recognition, Jian Zhao, Lin Xiong, Yu Cheng, Yi Cheng, Jianshu Li, Li Zhou, Yan Xu, Karlekar Jayashree, Sugiri Pranata, Shengmei Shen, Junliang Xing, Shuicheng Yan, Jiashi Feng, International Joint Conference on Artificial Intelligence (IJCAI), 2018.

[4] High Performance Large Scale Face Recognition with Multi-Cognition Softmax and Feature Retrieval, Yan Xu, Yu Cheng, Jian Zhao, Zhecan Wang, Lin Xiong, Karlekar Jayashree, Hajime Tamura, Tomoyuki Kagaya, Sugiri Pranata, Shengmei Shen, Jiashi Feng, Junliang Xing, IEEE International Conference on Computer Vision (ICCV) MS-Celeb-1M Workshop, 2017.

[5] Know You at One Glance: A Compact Vector Representation for Low-Shot Learning, Yu Cheng*, Jian Zhao*, Zhecan Wang, Yan Xu, Karlekar Jayashree, Shengmei Shen, Jiashi Feng, IEEE International Conference on Computer Vision (ICCV) MS-Celeb-1M Workshop, 2017. (* Indicates equal contributions.)

[6] Self-Supervised Neural Aggregation Networks for Human Parsing, Jian Zhao, Jianshu Li, Xuecheng Nie, Yunpeng Chen, Zhecan Wang, Shuicheng Yan, Jiashi Feng, IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Visual Understanding of Human in Crowd Scene Workshop, 2017.

[7] Multi-Human Parsing Machines, Jianshu Li*, Jian Zhao*, Yunpeng Chen, Sujoy Roy, Shuicheng Yan, Jiashi Feng, Terence Sim, International Journal of Computer Vision (IJCV) Under View, 2018. (* Indicates equal contributions.)

[8] Understanding Humans in Crowded Scenes: Deep Nested Adversarial Learning and A New Benchmark for Multi-Human Parsing, Jian Zhao, Jianshu Li, Yu Cheng, Li Zhou, Terence Sim, Shuicheng Yan, Jiashi Feng, ACM Multimedia (ACM MM), 2018 (Oral).

18-26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”26期“，获取直播地址。

特别鸣谢本次Webinar主要组织者：

VOOC责任委员：郭裕兰（国防科大）

活动参与方式：

1、VALSE Webinar活动依托在线直播平台进行，活动时讲者会上传PPT或共享屏幕，听众可以看到Slides，听到讲者的语音，并通过聊天功能与讲者交互；

2、为参加活动，请关注VALSE微信公众号：valse_wechat 或加入VALSE QQ群（目前A、B、C、D、E、F、G群已满，除讲者等嘉宾外，只能申请加入VALSE H群，群号：701662399）；

*注：申请加入VALSE QQ群时需验证姓名、单位和身份，缺一不可。入群后，请实名，姓名身份单位。身份：学校及科研单位人员T；企业研发I；博士D；硕士M。

3、在活动开始前5分钟左右，讲者会开启直播，听众点击直播链接即可参加活动，支持安装Windows系统的电脑、MAC电脑、手机等设备；

4、活动过程中，请不要说无关话语，以免影响活动正常进行；

5、活动过程中，如出现听不到或看不到视频等问题，建议退出再重新进入，一般都能解决问题；

6、建议务必在速度较快的网络上参加活动，优先采用有线网络连接；

7、VALSE微信公众号会在每周一推送上一周Webinar报告的总结及视频（经讲者允许后），每周四发布下一周Webinar报告的通知及直播链接。

[slides]