Geometric matching in image retrieval

In specific object image retrieval the goal is to find all images of a large collection that contain the same object as a user query image. When indexed images are also accompanied by information such as tags, then image retrieval methods can facilitate automatic content description of user images.

Geometric matching between images allows to detect a depicted object from different viewpoints and is a way to drastically increase the performance of image retrieval systems. This is however achieved at a higher computational cost, thus complexity of geometry based methods is crucial. The talk will focus on methods which are based on the Generalized Hough Transform and consider as input a set of correspondences established between points of a pair of images. We will present methods which start from local geometric constraints and are finally able to recover a geometric transformation between the two images. A different direction is to use weak and local geometric constraints that allow to estimate a total similarity score capturing the geometric consistency of the correspondences set without ever estimating an optimal transformation. At the end of the talk we will demonstrate an online retrieval engine which exploits such geometric matching methods to search in 2 million images.