At 10:00 on December 2, 2020, Professor Li Hongsheng (Chinese University of Hong Kong) made an online report entitled "Unsupervised and Domain Adaptive Object Re-identification" at the invitation of the School of Electrical and Automation Engineering, Nanjing Normal University. The report was held in the meeting room 202 of Qiming Building, and hosted by Professor Ding Shuye, Professor Min Fuhong, Zhang Lei, Xie Fei, Qian Weixing and other faculty members. More than 50 postgraduate students from the college participated in this academic report.
Professor Li Hongsheng received his PhD degree in Computer Science from Lehigh University, the United States in 2012. From 2013 to 2015, he was an associate professor at the University of Electronic Science and Technology of China. From 2015 to 2017, he served as a research assistant professor at the Chinese University of Hong Kong. Since 2018, he has been an assistant professor in the Multimedia Laboratory, the Chinese University of Hong Kong. His research interests include computer vision, deep learning and medical image processing. He is currently an associate editor of the international journal, Neurocomputing, and a guest editor of the International Journal of Computer Vision. He has published more than 70 papers in top-tier journals and conferences on computer vision, machine learning and medical image processing. He won the 2020 Outstanding Young Author Award of the IEEE Circuits and Systems Association and ranked the first place in the video object detection track of the ImageNet International Challenge in 2016.
Unsupervised learning and domain adaptation with the applications of object re-identification were introduced in this report. Professor Li Hongsheng shared the research works of “Mutual Mean-Teaching: Pseudo Label Refinery for Unsupervised Domain Adaptation on Person Re-identification”, “Self-paced Contrastive Learning with Hybrid Memory for Domain Adaptive Object Re-ID” and “Structured Domain Adaptation with Online Relation Regularization for Unsupervised Person Re-ID”, which were proposed by his research team. The introduced algorithms were published at top-tier machine learning conferences, i.e., International Conference on Learning Representations (ICLR) 2020 and Conference on Neural Information Processing Systems (NeurIPS) 2020. Object re-identification is at the core of smart city systems. In the era of deep learning, large-scale datasets promote the rapid development of this task. However, even a model trained on a large-scale dataset shows significant performance drops when directly being applied to a new camera system (or surveillance system). Moreover, the collection and manual annotation of training data is time-consuming and labor-intensive. The task of unsupervised and domain adaptive object re-identification is therefore proposed to tackle this challenge.
At the end of the report, the teachers and students passionately discussed with Professor Li Hongsheng regarding of the generalization abilities of supervised and unsupervised learning algorithms, the initialization and training protocols of coupled networks, video-related algorithms, etc. All the attendees benefited a lot from this report and had a better understanding on the tasks of unsupervised and domain adaptive object re-identification, as well as the researches in the field of machine learning, deep learning and computer vision.