Self-Paced Densely Connected Convolutional Neural Network for Visual Tracking

摘要：

　　Convolutional neural networks(CNNs)have achieved surprising results in visual tracking.To address the model drift problem,we propose a novel self-paced densely connected convolutional neural netwrok(SPDCT)to distinguish the reliable data from noisy and confusing data.In the proposed model,each sample is given a weight,which is estimated by SPDCT to indicate the reliability of the sample.The self-paced learning framework is then integrated into the online update phase to improve the robustness of CNNs.In order to determine the pace parameter of self-paced learning effectively,we propose an adaptive method based on the number of training samples.Meanwhile,with the aim of facilitating the representation power of the features,we enhance the feature reuse and the information flow by applying the densely connected learning.Extensive experimental results demonstrate competing performance of the proposed tracker over a number of state-of-the-art algorithms.

关键词： Model drift problem Self-paced learning Densely connected learning Convolutional neural networks

作者: Daohui Ge Jianfeng Song Yutao Qi Chongxiao Wang Qiguang Miao

作者单位: School of Computer Science and Technology,Xidian University,Xian 710071,Shaanxi,China;Xian Key Laboratory of Big Data and Intelligent Vision,Xian 710071,Shaanxi,China

会议类型: 国际会议

会议名称: 中国模式识别与计算机视觉大会(PRCV2018)

会议地点: 广州

会议语种:英文

页码: 103-114

在线出版日期: 2018-11-23（万方平台首次上网日期，不代表论文的发表时间）