Better Person Re-identification Using ResNet Model and Re-ranking Strategy

2018 2nd International Conference on Artificial Intelligence Applications and Technologies

计算机科学

Pan, Nengchao^1

Beijing Laboratory of Intelligent Information Technology, School of Computer Science, Beijing Institute of Technology, Beijing, China^1

关键词: Adaptive sparsities;    Computation complexity;    Person re identifications;    Re-ranking;    Transfer learning methods;   

Others  :  https://iopscience.iop.org/article/10.1088/1757-899X/435/1/012002/pdf
DOI  :  10.1088/1757-899X/435/1/012002

学科分类：计算机科学（综合）

来源: IOP

【 摘 要 】

In this paper, we propose a novel person re-identification algorithm based on deep learning. To start with, deep features are extracted from images by ResNet model. Then a set of related images are selected under the low similarity constraint. In the end, we re-rank the retrieval result with related images. Our algorithm has three advantages. First, the transfer learning method reduces the cost needed to train the model compared with the others. Second, we employ the genetic algorithm under adaptive sparsity constraint to detect related images, which reduces the computation complexity. Third, the detected related images can be applied to promote the initial retrieval result. Experiments on Market1501 have demonstrated the effectiveness of our algorithm and our algorithm is robust against variations in background, pose, illumination and viewpoint.