Abstract de la publi numéro 14353

In this paper, we address the problem of landmark image annotation, defined as the task of automatically annotating a landmark query image with relevant descriptors (keywords or tags). Given a new query image along with its geolocation metadata (latitude and longitude), we retrieve several other images already available in a community image database (e.g., flickr.com, panoramio.com), found within a fixed radius of the location of the query image. We then formulate the automatic landmark image annotation problem as a tag ranking problem over all the tags obtained from these pre-existing neighboring images. We propose several tag ranking factors, and by evaluating them against a gold standard constructed using the geolocation-oriented photo sharing platform panoramio.com, we show that an aggregated measure that combines both distance and frequency factors leads to results significantly better than any of the individual factors