Visual Localization at Scale

ABSTRACT

Visual localization refers to the recovery of the position and orientation of cameras in known or unknown scenes, using only visual data. We humans use vision as our primary source of information for localization, navigation, and exploration in our environments. Similarly, a high-quality image with a wide view of the surroundings often captures sufficient information to represent a location uniquely. Therefore, using images to localize agents in a map is an important research area of computer vision and robotics. Immediate consequences of solving the visual localization problem leads to many exciting and potentially revolutionary applications including; human assistance, autonomous systems, and augmented reality. Intuitively, by answering questions such as ``Where am I in the scene?” or “How to render objects for augmented reality?” In this talk, I will present the problem of visual localization in large scale scenes and recent progress in this direction. The talk contents will include, scene reconstruction, understanding, abstraction, and algebraic frameworks to exploit them for localization. Different applications of the visual localization systems will also be demonstrated.