Abstract:Pedestrian re-identification is a valuable problem that has enormous potential for practical applications like criminals monitoring and investigation,lost children search in the public area,personal photo album management,e-business,to name a few.Meanwhile,it is also very challenging since the appearance of the same person between camera views changes dramatically caused by illumination,viewpoint,pose,and background variations.Given a certain pedestrian feature space,because of the different discriminative ability of different feature components, learning a proper similarity is important for pedestrian re-identification.A dominant algorithm to learn a similarity is the metric learning that learns a Mahalanobis Similarity Function (MSF)to estimate the similarity of a pair of pedestrians.However,because MSF only projects a pair of pedestrians into the feature difference space and ignores the appearance of each individual,it is inadequate for modeling complex relationships of appearance variations from different cameras.In this paper,we propose to learn a Quadratic Similarity Function (QSF)that greatly strengthens the modeling ability of the similarity function.QSF is a generalization of MSF,and it not only represents the cross correlation relationship of a pair of pedestrians,but also describes the auto correlation relationship.So QSF is better to tackle large appearance variations of the same pedestrians than MSF.To learn a QSF,inspired from the classification perspective and the ranking perspective,we design two algorithms with different optimizing object functions.Experiments on the VIPeR dataset and the CUHK campus dataset show that both algorithms are effective to learn a QSF and their performance outperforms the previous algorithms.

Learning Quadratic Similarity Function for Pedestrian Re-Identification