Vision-based fine-grained location estimation

Abstract

In this chapter, we explore a variety of vision–based location estimation techniques, in which the goal is to determine the location of an image at a fine–grained level. First, we introduce the concept about image–based location and landmark recognition (Sect. 4.1), which determines the location of a given image by leveraging collections of geo–located images. Early techniques usually treat this as a similar image matching problem and use the geo–tags transferred from the matched database images. Some recent works have examined how to estimate more fine–grained and comprehensive geo–context information, such as the viewing direction estimation (Sect. 4.3) of photos. Next we will review the techniques for city–scale location recognition, informative codebook generation, and geo–visual clustering (Sect. 4.4). Moreover, we will introduce the structure–from-motion technique, which is closely related to estimating the camera geo–location by generating 3D models. With the 3D scenes reconstructed from the image collections, images are localized by 2D–3D alignment (Sect. 4.5). The camera location, viewing direction, and scene location are estimated simultaneously, which are essential to various applications. Moreover, another class of vision–based location estimation technique using satellite–imagery database is also described (Sect. 4.6).