Deep ensemble effectively and efficiently for vehicle instance retrieval

Abstract

This paper aims to highlight instance retrieval tasks centered around ‘vehicle’, due to its wide range of applications in surveillance scenario. Recently, image representations based on the convolutional neural network (CNN) have achieved significant success for visual recognition, including instance retrieval. However, many previous retrieval methods have not exploit the ensemble abilities of different models, which achieve limited accuracy since a certain kind of visual representation is not comprehensive. So we propose a Deep Ensemble Efficiently and Effectively (DEEE) framework, to preserve the impressive performance of deep representations and combine various deep architectures in a complementary way. It is demonstrated that a large improvement can be acquired with slight increase on computation. Finally, we evaluate the performance on two public vehicle datasets, VehicleID and VeRi, both outperforming state-of-the-art methods by a large margin.