I. INTRODUCTION
In recent years, research in the field of life-long navigation and localization of mobile robots has been considerably active [1], [2], [3]. Navigation systems used by mobile robots are typically based on Simultaneous Localization And Mapping (SLAM) approaches [4], [3]. In long-term SLAM implementations, the important task of detecting previously visited places is known as loop closure and is applied in the correction of map errors that accumulate over time [5]. In particular, systems whose main sensors are cameras typically solve loop closure by Visual Place Recognition (VPR) [6], [7], [8], [6], [9], [10], a fundamental and still open problem in computer vision.