Twitter Photo Geo-Localization Using Both Textual and Visual Features

摘要

In this paper, we propose a method to add geotags to Twitter photos which have no geotags. Our objective is localizing a Twitter photo using both textual features and visual features. For localization from texts, we use GeoNLP which estimates location from location names and location-dependent named entity. For localization from visual features, we use image search for a geotagged photo database. As visual features, we use local features and deep neural features and compare them. In the experiments, combining texts and deep neural feature based image search achieved the best results.