Improving Retrieval Efficiency of Person Re-Identification Based on Resnet50

Abstract

In recent years, the issue of person re-identification has become more and more popular, which is an important research subject in the field of computer vision, and many models or methods for different predicaments have been proposed successively. However, there are often differences between theory and practice. As a matter of fact, while collecting a large number of pedestrian images, retrieval efficiency becomes one of the significant evaluation indicators. Therefore, how to maintain high precision and quickly respond to retrieval requirements is a very important issue. This thesis explores many proposed person re-identification methods and improves retrieval time under the premise of maintaining a high precision rate. In this paper, we select Resnet50 as the feature output model, and use not only K-means Clustering to filter out the preliminary candidates but also Hierarchical Comparison to reduce the number of feature comparisons. The final experimental result shows the average retrieval time is improved dramatically with a speed-up ratio closed to 8, whereas the precision loss is under 3%.