News Resource: School of Optics and Photonics

Editor: News Agency of BIT

Translator: Guo Yating, News Agency of BIT

Recently, Xu Tingfa's research team from School of Optics and Photonics of Beijing Institute of Technology (BIT) has made new progress in the field of adversarial learning-related tracking. Related research results are titled as “Learning Context Restrained Correlation Tracking Filters via Adversarial Negative Instance Generation” , publishing in IEEE Transactions on Neural Networks and Learning Systems (IEEE TNNLS) (IF="10.451) , which is the top international journal in the field of artificial intelligence. IEEE" TNNLS is one of the most influential international academic journals in the field of artificial intelligence, ranking top among more than 140 JCR journals in this field in 2021, with an impact factor of 10.451, Q1. It mainly publishes and reports the latest research progress and technology in the fields of computational intelligence, artificial intelligence, data science and neural network. The first author of this work is Huang Bo, PhD of BIT, and the corresponding author is Professor Xu Tingfa of BIT.

The tracking performance of discriminant correlation filter(DCF) is often subject to unwanted boundary effects. In the past few years, many scholars have tried to solve the performance degradation caused by boundary effects through expanding the search area. However, introducing excessive background information makes the discriminative filter prone easy to learn from the surrounding context rather than the target. Prof. Xu Tingfa and his team first proposed a novel context restrained correlation tracking filter (CRCTF) to solve this scientific problem. The method suppressed the interference of similar background noise by adding high-quality adversarial generative negative instances. Specifically, the model constructs a generative adversarial GAN network at the initial frame to simulate the central target area using contextual background information. Aiming at the problem of slow computation speed of GAN network, the team proposed to use an efficient background motion vector estimation network to accelerate the generation of background instances in subsequent frames. Then, based on the generated background instances which the convolution suppression term is introduced, the original ridge regression objective equation is reformulated by the cyclic structure of correlation filtering and clipping operator. Finally, the alternating direction method of multipliers (ADMM) is used to solve the tracking filter at high speed in frequency domain.

Figure 1 Adversarial background instance generation model

This study explores the influence of adversarial instances to optimize model learning on the training of ridge regression correlation filters. The model uses high quality antagonism to generate negative instances and has excellent performance in suppressing context-like noise. CRCTF shows significant performance advantages over the most advanced and highly optimized baseline correlation filtering algorithm on multiple challenging tracking data sets. The effectiveness of adversarial learning for optimizing ridge regression trackers is demonstrated through qualitative and quantitative analysis on benchmark evaluation indexes. This research is helpful to understand the inner relationship between the regular terms of the ridge regression target equation more clearly and deeply, which is of great significance to the development of the correlation filtering target tracking field.

Paper details: Bo Huang, Tingfa Xu, Jianan Li, Fei Luo, Qingwang Qin, Junjie Chen, "Learning Context Restrained Correlation Tracking Filters via Adversarial Negative Instance Generation[J]", IEEE Transactions on Neural Networks and Learning Systems, 2021, doi: 10.1109/TNNLS.2021.3133441.

Paper link:https://ieeexplore.ieee.org/document/9662066

About the first author:

Huang Bo, a 2016 master-doctoral PhD of School of Optics and Photonics of BIT, majoring in computer vision and deep learning under the tutelage of Professor Xu Tingfa. He has published 20 academic papers, 11 of which have been published as the first author in IEEE TNNLS, IEEE TCYB, IEEE TMM, PR, Neurocomputing and other high-level journals or conferences, with an impact factor of 49.349. He has applied for 3 Chinese invention patents and 3 soft papers. At the same time, he has been the reviewer of IEEE TCSVT, Neurocomputing, IEEE Access, CVPR and other SCI journals or conferences for many times. He won the ICCV 2021 Anti-UAV Workshop & Challenge champion and the best paper award, second prize in Photoelectric Design Competition, the first prize in “Challenge Cup” Capital Competition, third prize in National Post-Graduate Mathematical Contest in Modeling, grand prize of “Century Cup”, and third prize of "encyclopedia" cup of Electronic Design Contest, second prize of Light: Science & Applications Beijing, National Scholarship, Beijing Outstanding Graduates, Beijing Institute of Technology Outstanding Doctoral Dissertation Nurturing Fund and other honors.

About the corresponding author:

Xu Tingfa, professor, PhD supervisor, professorial chair of National Key Disciplines “Optical Engineering”, deputy director of Key Laboratory of Photoelectronic Imaging Technology and System (Beijing Institute of Technology), Ministry of Education. In recent years, he has led his research team to deepen research on photoelectric imaging detection and recognition, hyperspectral imaging processing and other directions. He has presided over more than 30 major scientific instrument development projects of the National Natural Science Foundation of China. He has published more than 120 academic papers in a series of international and domestic journals, including more than 80 papers indexed by SCI/EI. As the first inventor, he has applied for 40 national invention patents, 15 of which have been authorized or publicized.