Modeling Unknown Class Centers for Metric Learning on Person Re-Identification

Metric learning is one of the major methods for person re-identification. Most existing metric learning methods for person re-identification first generate the pairwise constraints where the sample pairs with the same labels consist of the positive set and the ones with the different labels comprise the negative set. And then, different kinds of methods are formulated to pull the pairs in positive set together and penalize the pairs in negative set such that a good metric is learned. However, such a process has two drawbacks: 1) the size of negative set is often far more than the negative set for which the learning process is largely dominated by the large amount of negative sample pairs and 2) it often experiences tedious optimization procedures to compute pairwise distances which would be computationally intractable in real scenarios, especially for large-scale data sets. To address the above issues, we propose a new simple and effective metric learning method, which gets rid of the pairwise constraints. We take the unknown class center's information into consideration to model the relationships between different classes directly, and a new objective function is formed. In addition, the proposed objective function can be solved by taking advantage of simple matrix multiplications and hence can avoid computationally complex optimization schemes. The proposed algorithm termed largest center-specific margin metric learning is shown to be computationally efficient and can be applied to large-scale person re-identification. Extensive experiments carried out on two challenging large-scale databases (CUHK03 and Market1501) demonstrate that the proposed algorithm performs favorably against the state-of-the-art approaches.