Mask Multi-Head Attention with Partition Network for Vehicle Re-Identification

Vehicle Re-identification aims to match a specific vehicle image across different places or cameras based on the similarity among vehicles. vehicle re-id remains confronted with two severe challenges, small inter-class variability caused by a similar vehicle with a similar type and color, and dramatic intra-class variability caused by the variation of view. More recently, methods are proposed to improve performance by using additional metadata such as critical points and orientation, which all require expensive annotations. Therefore, we introduce attention mechanism to solve these two problems without considering extra annotation. In this paper, we propose a novel mask multi-head attention with partition network (MMAPN). To discover subtle differences between two similar vehicles, we propose a partition unit to discover more local detail. To extract features that are robust to both tremendous intra-class differences and subtle inter-class variability, we propose a mask multi-head attention block to extract potential features. Extensive experimental evaluations show our approach achieved state-of-the-art performance.