Smart Features for Dynamic Vision Sensors

This paper presents a semi-supervised procedure for training a fully convolutional neural network for interest point detection and description. Contrary to previously trained networks utilized for the same purpose, our network was tailored to work with event-based imagery from a downward facing camera aboard a fixed-wing unmanned aerial vehicle. Event-based cameras are a novel type of visual sensor that operate under a unique paradigm, providing asynchronous data on the log-level changes in light intensity for individual pixels. This hardware-level approach to change detection allows these cameras to achieve ultra-wide dynamic range and high temporal resolution. The final system produces state-of-the-art repeatability and homography estimation results on an aerial event-based image dataset when compared to traditional interest point detector and descriptor algorithms.