SVDNetforPedestrianRetrieval.PDF

SVDNet for Pedestrian Retrieval Yifan Sun ,Liang Zheng ,Weijian Deng ,Shengjin Wang Tsinghua University University of Technology Sydney University of Chinese Academy of Sciences sunyf15@mails.tsinghua.edu.cn 7 1 0 2 Abstract r p A This paper proposes the SVDNet for retrieval problems, with focus on the application of person re-identification (re- 0 ID). We view each weight vector within a fully connected 2 (FC) layer in a convolutional neuron network (CNN) as a ] projection basis. It is observed that the weight vectors are V usually highly correlated. This problem leads to correla- C tions among entries of the FC descriptor, and compromises . the retrieval performance based on the Euclidean distance. s c To address the problem, this paper proposes to optimize the [ deep representation learning process with Singular Vector Decomposition (SVD). Specifically, with the restraint and Figure 1: A cartoon illustration of the correlation among 3 v relaxation iteration (RRI) training scheme, we are able to weight vectors and its negative effect. The weight vectors 3 iteratively integrate the orthogonality constraint in CNN are contained in the last fully connected layer, e.g., FC8 9 training, yielding the so-called SVDNet. We conduct ex- layer of CaffeNet [ 12] or FC layer of ResNet-50 [11]. There 6 periments on the Market-1501, CUHK03, and DukeMTMC- are three training IDs in red, pink and blue clothe