Pixels to Location: Broadening Image Place Discovery Through Low-Resolution Satellite Image Search

Today we increasingly rely on digital maps to discover un-familiar locations. The primary mechanism utilized by most digital map applications depends on GPS signals. However, these signals can be obstructed or weakened by tall buildings, thereby impacting their reliability in urban areas. To enhance the accuracy of digital maps, the concept of image geo-localization is gaining traction. Among these approaches is cross-view image geo-localization, where images needing location identification are cross-referenced with a database of aerial images tagged with geographic information. One of the main hurdles in implementing such techniques on a larger scale is the demand for high-resolution geo-tagged aerial image databases. This limitation restricts its practical applicability due to increased cost in procuring high resolution aerial imagery. In this work we present pixels to location, a framework that enables low resolution aerial images in the search data-base. Experimental results demonstrate that our approach can achieve a top-1% recall rate of 92.33% utilizing aerial view resolution of 50 cm / pixel making it suitable for practical applications that necessitate a broader search area.