Multi-branch Fusion Fully Convolutional Network for Person Re-Identification

Abstract

Building effective CNN architectures with light weight has become an increasing application demand for person re-identification (Re-ID) tasks. However, most of the existing methods adopt large CNN models as baseline, which is complicated and inefficient. In this paper, we propose an efficient and effective CNN architecture named Multi-branch Fusion Fully Convolutional Network (MBF-FCN). Firstly, multi-branch feature extractor module focusing on different receptive field sizes is designed to extract low-level features. Secondly, basic convolution block units (CBU) are used for constructing candidate network module to obtain deep-layer feature presentation. Finally, head structures consisted of multi-branches will be adopted, combining not only global and local features but also lower-level and higher-level features with fully convolutional layer. Experiments demonstrate our superior trade-off among model size, speed, computation, and accuracy. Specifically, our model trained from scratch, only has 2.1 million parameters, 0.84 GFLOPs and 384-dimensional features, reaching the state-of-the-art result on Market-1501 and DuckMTMCreID dataset of Rank-1/mAP = 94.5$\%$/84.3$\%$, Rank-1/mAP = 86.6$\%$/73.5$\%$ without re-ranking, respectively.