Partial person re-identification with two-stream network and reconstruction

Abstract

Partial person re-identification is a challenging issue at present. However, affected by occlusions, features in person re-identification cannot be detected and the traditional person re-identification methods can not accurately deal with it. In order to solve this problem, we propose to match query and gallery by combining different modes from two-stream network with sparse reconstruction to realize partial person re-identification. For acquiring features, bilinear pooling is applied to fuse the two different modes from the appearance network and pose network aiming at better performance. For matching query and galley, the robust sparse representation reconstructs the features extracted by the network for flexible solution, using the parameters learned from galley. The reconstruction process achieves arbitrary size images in partial person re-identification. In addition, we extract mid-level feature and fuse it with the high-level feature for more accuracy. Experiments demonstrate the performance of the proposed method better compared with the methods of state-of-the-art person re-identification methods on dataset Market1501, CUHK03, DukeMTMC-reID and partial person dataset Partial-REID, Partial-iLIDS.