I. INTRODUCTION
With taking images being rather cheap people typically take
many pictures of the same scene to maximize the possibility of
getting a good picture. In such sets one typically finds
overlapping images, like one picture of the full Eiffel Tower and
a second one showing just the tip of the tower in detail. In this
example the mentioned two images feature an explicit pair wise
spatial relation. The picture showing the top of the Eiffel tower
captures a detail in the upper part of the other image. Another
example is shown in Figure 1. The leftmost image shows the boat
in detail, the image in the center shows the horizon and the
sunset, while the third image captures the scene in a broader
context. In many scenarios this overlap in content and the explicit
spatial is relevant to users trying to find and retrieve images. In
other words: if an image A is found by a search engine and has a
spatial relation with the images B, C and D, these images have a
high possibility of being relevant for the user as well.
While we are aware of the Photosynth project [11], it turned
out that current search engines do not provide a feature to browse
or retrieve images based on explicit spatial relations. No current
search engine or result visualization allowed for retrieving
images that were “left” or “right” of a query image or showed a
detail in a certain region to propose potentially interesting images
to the user. Therefore, we focus in our work on a simple and
robust method to (i) identify overlap between photos and (ii)
classify the overlaps to find and name the spatial relation with a
newly proposed taxonomy to allow easy and fast indexing and
search. We do not investigate advanced methods for image
registration [7] or image stitching [8], but focus on the retrieval
aspect of spatial relations between images. The approach and the
application described in this paper features an algorithm, which is
fast and robust and its results can be used to create an inverted
list to improve current image search engines.
Our work addresses the problem of (i) how to identify
overlapping image contents, (ii) how to identify explicit spatial
relations between different images based on the found overlap,
and (iii) how to create a taxonomy to enable indexing of spatial
relations for browsing and search. Note that the proposed method
does not index image content but identifies and indexes the
existing spatial relations between two images to improve current
search results.