|M.Sc Student||Sezganov Dmitry|
|Subject||Large Scale Content-Based Image Retrieval Using Geometric|
|Department||Department of Electrical Engineering||Supervisor||Professor Moshe Porat|
|Full Thesis text|
Visual search in very large image collections has become a timely research problem. We focus on specific object recognition search in which the goal is to find all the occurrences of a particular object in a very large image database, despite potential changes in scale, location and viewpoint, as well as changes in illumination, background clutter and partial occlusion. Typically, the images of interest (e.g., photos of landmarks) are indexed in a database enabling later search by visual queries. Such a database can contain millions of images, posing challenges in terms of memory, computational requirements and recognition performance. In this work, we concentrate on exploiting geometric information in these methods. We focus on the computational requirements in order to make them suitable for real-time applications.
The bag-of-words (BOW) model is commonly used in information retrieval. Originally developed for text, it has been adapted for visual search. In spite of similarity, there are different challenges. Less effective weighting schemes and also quantization errors reduce the discrimination power of a single visual word. The standard BOW model completely ignores geometric relationship information between visual words. Geometrical information is usually involved only in the post-processing spatial verification step usually implemented with the RANdom SAmple Consensus (RANSAC) algorithm. To enable visual search in real-time, however, RANSAC can be applied only to a relatively small number of top candidates due to its computational requirements. We propose two new efficient algorithms for spatial verification between two sets of local features that can be integrated with inverted files. These algorithms significantly improve the initial ranking of the search results, promoting suitable candidate images to the top of the list. Experimental results show that the proposed methods outperform the baseline BOW followed by full geometric verification based on RANSAC. Full geometric verification can be complementary to further improve the search results.