Improved Local Maximal Occurrence Feature Representation for Person Re-Identification

Person re-identification is a technique to automatically search the same people in different camera views. It is a challenging task because of the huge variations caused by different illuminations and views in images. There is an effective method tackling the problem, which learns a useful feature representation for the pedestrian images. In recent research, a feature representation named Local Maximal Occurrence (LOMO)achieves outstanding performance, but LOMO has a disadvantage that its robustness to illumination is not excellent. In this paper, we propose an improved LOMO representation to further reduce the impact of light. The improved LOMO feature combines Color Name feature with LOMO, because Color Name feature has strong robustness to illumination. However, Color Narne feature has the weakness that it loses graphical details. Therefore, the feature is unsuitable for traditional pooling method. To overcome the disadvantage, a specific pooling method called global Gaussian pooling for Color Name feature is designed, which is based on Gaussian distribution and can weigh the different regions of an image. The important regions gain more attention in matching images. The experimental results on two public available datasets demonstrate the effectiveness of our proposed approach.