We choose to classify annotations into two dierent categories:
scene-based and object-based (See Figure 5). Scene
annotations are like image captions where we store information
related to the buildings or structures present.
Object annotations are where a particular structure, building,
or an interesting architecture is marked by a rectangular
boundary on the image and a description is stored for
it. When we nd such annotations in the annotation source
(the best match for a query), we mark the object's boundary
on the query image. The object boundary information
is stored in the form of four corner points representing a
rectangular region in the image. In order to transfer the
object boundary from the source image, we estimate a homography
with the query image. For this, we use the inlier
matches computed during the spatial verication step.