Towards Good Practices for Image Retrieval Based on CNN Features

Abstract

Recent works have demonstrated that Convolutional Neural Networks (CNNs) achieve state-of-the-art results in several computer vision tasks. CNNs have also shown their ability to provide effective descriptors for image retrieval. In this paper, we focus on CNN feature extraction for instance-level image search. We started by studying in depth several methods proposed to improve the Regional Maximal Activation (RMAC) approach. Then, we selected some of these advances and introduced a new approach that combines multi-scale and multi-layer feature extraction with feature selection. We also propose an approach for local RMAC descriptor extraction based on class activation maps. Our parameter-free approach provides short descriptors and achieves state-of-the-art performance without the need of CNN finetuning or additional data in any way. In order to demonstrate the effectiveness of our approach, we conducted extensive experiments on four well known instance-level image retrieval benchmarks (the INRIA Holidays dataset, the University of Kentucky Benchmark, Oxford5k and Paris6k).