| Abstract: |  The thesis presents the location recognition method of a query image in a dataset of images labelled with geo-location information. This is a challenging task as the imaged appearance of the query image can be very different from the appearance of images in the database due to changes in viewpoint, lighting, and partial occlusion by other objects. In addition, the query image might be captured at a different time of the day, different season or in a different year altogether.
We employ for this task the bag-of-visual-words approach with large vocabularies and fast spatial matching previously used for object retrieval in large image collections. First, the thesis reviews retrieval approaches and their relation to the localization problem. Second, we discuss a method to create the visual vocabulary using the geolocation of database images. Third, query expansion by a non-geotagged image database (such as Panoramio or Flickr) as a way to enrich the query image and a significant improvement of image enriching on extremely hard query images is presented. Fourth, an approach to improving the localization using the automatic detection and suppression of confusing and non-informative features in the geotagged database was developed. Finally, a localization of an image sequence (video) based on Bayes filtering is presented and an improvement of the localization is shown.
We experimentally evaluated image based localization performance and localization stages benefits on several real city-streets different datasets of almost 20K images. We also present the effect of choosing different detector/descriptor types and vocabulary sizes.
|
|---|