Recently, at the international computer vision summit CVPR (IEEE Conference on computer vision and pattern recognition), Professor luhuchuan's team achievements of the school of information and communication engineering of the Department of electronic information and electrical engineering of our University won the nomination for the best paper of cvpr2020.
Computer vision is the hottest research field of artificial intelligence. CVPR is the most influential conference in this field. It can be seen from the Google academic influence ranking that the conference ranks 10th (nature and science rank 1st and 3rd respectively).
A total of 6656 submissions were received in this CVPR, 1470 of which were accepted, with an acceptance rate of 22.09%, of which 26 were nominated for the best papers, with an acceptance rate of only 0.39%. The first author of this achievement is Dai Kenan, a master of the school of communication and communications of our university. The instructors are Wang Dong, lijianhua and luhuchuan. In addition, a total of 8 papers of iiau laboratory led by Professor luhuchuan were hired by cvpr2020 this year, and other instructors include zhanglihe, piaoyongri, etc.
In addition, the iiau team won three championships in vot2020, the most authoritative international competition for target tracking! There are five tracks in this VOT competition, among which the long-term track, real-time track and deep track champions were won by Dai Kenan, Yan Bin and Wang Yingming, our master students respectively!
This is the fourth consecutive year that the iiau team has won the championship in VOT - vot2019 won the long-term track championship by master Dai Kenan, vot2018 won the long-term track championship by Master Zhang Yunhua, and vot2017 won the first place in the open group by Dr. Sun Chong. High performance long term tracking with meta Updater.
In recent years, long-term tracking has attracted more and more attention because it is closer to practical applications. In the long-term tracking, because the video is very long and there are a lot of challenges such as disappearing out of the country, online updates are full of risks, which also leads to the poor performance of many short-term tracking SOTA algorithms in the long-term tracking. In this paper, a long-term update controller is proposed, which encapsulates the geometry information, discrimination information and appearance information obtained by on-line tracking and sends them to the long-term and short-term memory network, and then makes a binary classification to judge whether the current tracking state can be updated.
In addition, this paper also proposes a long-term tracking framework, which is composed of a short-term tracker, an update controller, a full graph detector and a verifier. The short-term tracker is used for local tracking. When the target is lost, the full graph detector is used to detect the candidate target, the verifier judges, and the update controller controls the update of the short-term tracker and the verifier. Each module is relatively independent. This scheme makes the performance of long-term tracking better benefit from the development of short-term tracker and full image detector.