Preserving Structural Relationships for Person Re-Identification

Recent processes on many computer vision and multimedia researches heavily rely on Convolutional Neural Network (CNN) with pooling layer incorporated, where pooling operation reduces the amount of parameters and brings in translation invariance. However, we discover that pooling operation may destroy valuable structural relationship information, leading to defective feature learning in tasks such as person re-identification. In this paper, we propose a method called Structural Relationship Learning (SRL) to capture structural relationships by constructing a spatially structured graph based on the convolved features and propagate information over the edges. Coupled with pooling operation by metric fusion, SRL provides more comprehensive information for identity discrimination. Experiments are conducted on the iLDIS-VID, PRID2011 and MARS datasets and the results demonstrate the advantages of our proposed method.