最新消息

2016年10月19日学术报告:Aggregated Deep Feature Representation for Visual Search


添加时间:2016-10-17 12:20:10



题目:Aggregated Deep Feature Representation for Visual Search

时间:2016年10月19日(周三),下午1点30分

报告人:Yuwei Wu

地点:研究生楼203

 

Abstract

Several recent works have shown that the activations from the convolutional layers of deep convolutional neural networks (CNNs) can be interpreted as local features describing particular image regions. However, how to effectively exploit these convolutional features for visual search is still a challenging problem. Most existing aggregation schemes of convolutional features only compute first-order statistics, e.g., average pooling and max pooling, and thus neglect the relationships among different feature maps. In this talk, we advocates exploiting appropriately convolutional layer activations to constitute a powerful descriptor which is achieved by covariance matrix representation and nonlinear kernel matrix representation. More specifically, the covariance matrix captures the second-order statistics of convolutional features to produce a compact and powerful global descriptor. Since covariance matrix only evaluates linear correlation of features, we further introduce a kernel matrix as a global descriptor which models a nonlinear relationship among deep convolutional features. These two global representations lie in the space of symmetric positive definite (SPD). Through the theory of non-Euclidean Riemannian geometry, SPD matrices often turn out to be better suited in capturing several desirable data properties. Therefore, the resulting representation thus can encode high level variation information among deep convolutional features. Comprehensive experiments on the challenging Oxford5k, Sculpture6k, Paris6k, Holidays, Holidays+100K and UK Bench datasets show that the new compact global descriptor notably outperforms the state-of-the-art methods.

Yuwei Wu received the Ph.D. degree in computer science from Beijing Institute of Technology (BIT), Beijing, China, in 2014. He is a post doctoral research fellow at the school of Electrical & Electronic Engineering, Nanyang Technological University, Singapore from August 2014. He has strong research interests in computer vision. He received Distinguished Dissertation Award Nominee from China Association for Artificial Intelligence (CAAI), Academic Scholarship for Ph.D. Candidates from Ministry of Education, P.R.China, and Distinguished Ph.D. Thesis award and XU TELI Excellent Scholarship from BIT.



作者:宋浩