Person re-identification by bidirectional projection

Person re-identification plays an important role in video surveillance system. It can be regarded as an image retrieval process which aims to find the same person in multi-camera networks. Many existing methods learn a pairwise similarity measure by mapping the raw feature to a latent subspace to make the data more discriminative. However, most of these methods project all the data into the same subspace ignoring the different characteristics that the same person and different person hold. To solve the aforementioned problem, a pairwise based method is proposed by projecting the raw feature onto two discriminative subspaces according to whether a image pair is of the same class. The proposed method constructs a relative and pairwise model by using the logistic loss function to give a soft measure of the pairwise loss. Meanwhile, a trace norm regularization is used to create the convexity of the objective function, which also help to limit the dimension of the subspaces. Experiments carried on the benchmark dataset VIPeR show that the proposed model obtains better results compared with state-of-the-art methods.