Abstract
Differential Visual Streams (DiVS) is a method to recognize common locations and quantify viewpoint differences between two image sequences, enabling robot self-localization along known paths (for which reference images exist) from images recorded by a single uncalibrated camera. It combines concepts from memory networks and convolutional networks, recent developments in machine learning with complementary features. Experiments show DiVS provides a sound basis for an image-based teach-replay navigation system working both indoors and outdoors, against variations in lighting and landscape composition.
