Bi-Directional Style Adaptation Network for Person Re-Identification

Abstract

Large-scale surveillance camera system can provide multiple visual information. Nevertheless, the images exhibit diverse appearance features due to different parameters and installation positions of the cameras. This camera style variation deteriorates its benefit from capturing identity features in person re-identification (Re-Id). The existing methods for filtering shallow appearance information through the Instance Normalization (IN) layer are extremely unfavorable to supervised tasks. To mitigate this problem, a simplified and valid Bidirectional Style Adaptation Network (BSA-Net) is presented to incorporate a new branch containing IN layer to learn invariant features with changing appearance. For these two branches, the structure is completely independent and the parameters are partially shared. BAS-Net is able to focus on the extraction of identity information while preserving appearance features. Specially, this new branch is removed during the testing phase, which significantly facilitates performance without introducing new computation. The superiority of the model is confirmed in extensive experiments on widely used benchmarks.