您的位置:首页  > 论文页面

一种基于CNN特征的域适应Egocentric视频有意义帧提取算法

发表时间:2018-11-30  浏览量:1804  下载量:166
全部作者: 田贺英,尹辉,杨林,欧伟奇
作者单位: 北京交通大学计算机与信息技术学院,交通数据分析与挖掘北京市重点实验室
摘 要: Egocentric视频数据量大且信息冗余度高,有意义帧的提取对于Egocentric视频的分析和理解具有重要意义。本文提出一种基于卷积神经网络(convolutional neural network,CNN)特征的域适应Egocentric视频有意义帧提取算法,该算法避免了构建Egocentric视频人工标注训练数据的巨大工作量,以已标注有意义帧的固定视角视频数据为源样本,分别提取基于CNN的源域和目标域(Egocentric视频)特征空间,构建连接源域和目标域的域不变特征空间,采用基于测地线流内核的方法计算源样本和目标样本的相似度,根据相似度度量提取Egocentric视频中的有意义帧。通过在UT Egocentric数据集上的对比实验发现,在有意义帧提取的准确度上,本文算法较Web Prior+DA等算法提升了至少4个百分点,验证了本文提出算法的有效性。
关 键 词: 计算机应用;Egocentric视频;卷积神经网络;特征选择;域适应;有意义帧
Title: A domain-adapted algorithm for detecting meaningful frames in Egocentric video based on CNN feature
Author: TIAN Heying, YIN Hui, YANG Lin, OU Weiqi
Organization: Beijing Key Lab of Traffic Data Analysis and Mining, School of Computer and Information Technology, Beijing Jiaotong University
Abstract: Egocentric video is characterized by large amounts of data, containing too much redundant data, so the extraction of meaningful frames is of great significance to the analysis and understanding of Egocentric video. This paper proposed a domain-adapted algorithm for detecting meaningful frames in Egocentric video based on convolutional neural network (CNN) feature. The algorithm uses fixed view video data which has been labeled with meaningful frame as source domain, extracting the feature space of the source domain and the target domain (Egocentric video) based on CNN feature separately, which avoids the tremendous workload of building Egocentric video manually annotation training data. Then the domain invariant feature space connecting the source and target domains is constructed. The similarity between the source domain and the target domain is calculated based on geodesic stream kernel method. Finally, meaningful frames in Egocentric video are extracted according to similarity measures. Using contrast experiment on UT Egocentric data set, our algorithm has improved by at least 4% over other algorithms on the accuracy of meaningful frame extraction, such as Web Prior+DA, which validates the effectiveness of our algorithm.
Key words: computer applications; Egocentric video; convolution neural network (CNN); feature extraction; domain-adapted; meaningful frames
发表期数: 2018年11月第22期
引用格式: 田贺英,尹辉,杨林,等. 一种基于CNN特征的域适应Egocentric视频有意义帧提取算法[J]. 中国科技论文在线精品论文,2018,11(22):2234-2244.
 
7 评论数 0
暂无评论
友情链接