Efficient Multi-Granularity Network based on Local Context-aware Correlation Feature for Person Re-Identification

摘要

The joint learning strategy of the global and local feature can further improve discriminative ability in complex realistic scenarios. The adjacent horizontal stripes regions of local features have the strong correlation characteristics. Considering the structure characteristics of human body, in this paper we design the Multi-Granularity Network based on Local Contextaware Correlation Feature (MGN_CACF) based on ResNet50-IBN-a backbone, which is split into four branches including a Top DropBlock branch, a global feature branch and two local feature branches. Experimental results shown that our proposed algorithm can achieve significant performance improvement, such as for Market-1501 and CUHK03 datasets Rank-l/mAP = 95.7%/88.5% and Rank-l/mAP=78.0%/73.8% respectively.