Local region partition for person re-identification

Abstract

Due to the different posture and view angle, the image will appear some objects that do not exist in another image of the same person captured by another camera. The region covered by new items adversely improved the difficulty of person re-identification. Therefore, we named these regions as Damaged Region (DR). To overcome the influence of DR, we propose a new way to extract feature based on the local region that divides both in the horizontal and vertical directions. Before splitting the image, we enlarge it with direction to increase the useful information, potentially reducing the impact of different viewing angles. Then each divided region is a separated part, and the results of the adjacent regions will be compared. As a result the region that gets a higher score is selected as the valid one, and which gets the lower score caused by pose variation and items occlusion will be invalid. Extensive experiments carried out on three person re-identification benchmarks, including VIPeR, PRID2011, CUHK01, clearly show the significant and consistent improvements over the state-of-the-art methods.