IEEE International Conference on Multimedia Expo (ICME) 2018 was held in San Diego, US on July 23rd-27th, 2018. One VIPL research paper, "Shen Li, Bingpeng Ma, Hong Chang, Shiguang Shan, Xilin Chen. Continuity-Discrimination Convolutional Neural Network for Visual Object Tracking", was awarded as one of the Platinum Best Papers. This year, ICME reviewers selected one Diamond Best Paper and six Platinum Best Papers to acknowledge these works in terms of novelty and research value both in theory and practice.
This VIPL research paper introduces temporal slowness into visual tracking, which is regarded as a potential learning principle in human visual cells. The authors propose temporal appearance continuity to characterize the learning principle and mathematically prove that the resulting representation error can be thereby sufficiently small with high probability. Besides, a novel notion "object-centroid" is proposed to alleviate drifting and inaccurate localization. These two notions are formalized as losses that supervise the learning process of CDCNN. Ablation studies validate the contributions of this research paper and experimental results demonstrate the competitive performance of the proposed method against state-of-the-art trackers as shown below.
Download: