Uploading tourist photographs is a popular activity on photo sharing platforms. The manual annotation of these images is a tedious process and the users often upload their images with no associated textual information. Automating the annotation process has received a lot of attention but the problem remains a hard one, especially when dealing with large and heterogeneous databases. Here we focus on landmarks images, very frequent among tourism pictures, and propose a new automatic technique for annotating this type of pictures. Our system, called MonuAnno, relies on the joint exploitation of localization information and of image content analysis in an efficient and scalable framework. The annotation is performed using a two steps k Nearest Neighbors (k-NN). First, only neighboring landmarks of a new unlabeled georeferenced image will be considered as potential annotations and the image will be attributed to the landmark that is visually closest. Then, we introduce a verification step that eliminates false positives (images taken near a landmark that represent something else). The technique was tested on Web images and the results show that the precision of the labeling process in MonuAnno exceeds 80%, when annotating around 50% of the images in the test set.
Mendeley saves you time finding and organizing research
Choose a citation style from the tabs below