Landmark Recognition: From Small-Scale to Large-Scale Retrieval

Abstract

During the last years, the problem of landmark recognition is addressed in many different ways. Landmark recognition is related to finding the most similar images to a starting one in a particular dataset of buildings or places. This chapter explains the most used techniques for solving the problem of landmark recognition, with a specific focus on techniques based on deep learning. Firstly, the focus is on the classical approaches for the creation of descriptors used in the content-based image retrieval task. Secondly, the deep learning approach that has shown overwhelming improvements in many tasks of computer vision, is presented. A particular attention is put on the major recent breakthroughs in Content-Based Image Retrieval (CBIR), the first one is transfer learning which improves the feature representation and therefore accuracy of the retrieval system. The second one is the fine-tuning technique, that allows to highly improve the performance of the retrieval system, is presented. Finally, the chapter exposes the techniques for large-scale retrieval, in which datasets contain at least a million images.