Person Re-identification with Discriminative Dictionary Learning

Person re-identification (re-id) is important for video surveillance, and has interestingly algorithm challenges and extensive practical applications. Recently, the Sparse Representation based Classification (SRC) produced excellent results in person re-identification, in which the Dictionary Learning (DL) method is a very important part. Discriminative power of the learned dictionary determines the performance of re-identification. Previous approaches usually discriminatively train the dictionary by enforcing explicit constraints on DL. In this paper, we propose a heuristic discriminative dictionary learning method which improves the discriminative power of dictionary by transforming the representation space of dictionary in training. First, we figure out the statistical distribution of the training data and divide the data into two categories. Second, a transformation function is used to change the dictionary's expression space. The dictionary learned by our method is proved to be effective in person re-id. Experiments on the benchmark dataset (CAVIAR4REID, i-LIDS) demonstrate that the proposed method outperforms the state-of-the-art approaches.