A Network with Feature Partition and Local Feature Embedded for Vehicle Re-Identification

Vehicle re-identification is a sophisticated computer vision task. The existing methods mainly focus on global feature processing, which find it difficult to address the inter-class similarity and the intra-class difference. This paper proposed a network with detail-oriented ability and contextual information reunion, which integrated global features, local features and partition features into a unified framework for exploring more details to re-identify the vehicles. The network utilized a local branch with the local self-attention module (LSA) to explore the details within the marked regions, which can identify and distinguish the subtle differences from the local regions of different vehicles. Also, the network used the partition branch to explore the common contextual information between different parts for clustering images of vehicles with the same id. The experiments on VeRi776 datasets verified the effectiveness of the proposed method.