Towards real-time interest point detection and description for mobile and robotic devices

Convolutional Neural Networks (CNNs) have been successfully adopted by state-of-the-art feature point detection and description networks for the past number of years. The focus of these systems has been predominately on the accuracy of the system, rather than on its efficiency or ability to be implemented in real-time on embedded robotic devices. This paper demonstrates how techniques, developed for other CNN use cases, can be integrated into interest point detection and description systems to compress their network size and reduce the computational complexity; this reduces the barrier to their uptake in computationally challenged environments. This paper documents the integration of these techniques into the popular Reliable Detector and Descriptor (R2D2) network. Along with the integration details, a comprehensive Key Performance Indicator (KPI) framework is developed to test all aspects of the networks. As a result, this paper presents a lightweight variant of the R2D2 network that significantly reduces parameters and computational complexity while crucially maintaining an acceptable level of accuracy. Consequently, this new compressed network is more appropriate for use in real world systems and advances the efforts to implement such CNN based system for mobile devices.