Efficient object recognition and image retrieval for large-scale applications

Abstract

Algorithms for recognition and retrieval tasks generally call for both speed and accuracy. When scaling up to very large applications, however, we encounter additional significant requirements: adaptability and scalability. In many real-world systems, large numbers of images are constantly added to the database, requiring the algorithm to quickly tune itself to recent trends so it can serve queries more effectively. Moreover, the systems need to be able to meet the demands of simultaneous queries from many users. In this thesis, I describe two new algorithms intended to meet these requirements and give an extensive experimental evaluation for both. The first algorithm constructs an adaptive vocabulary forest, which is an efficient image-database model that grows and shrinks as needed while adapting its structure to tune itself to recent trends. The second algorithm is a method for efficiently performing classification tasks by comparing query images to only a fixed number of training examples, regardless of the size of the image database. These two methods can be combined to create a fast, adaptable, and scalable vision system suitable for large-scale applications. I also introduce LIBPMK, a fast implementation of common computer vision processing pipelines such as that of the pyramid match kernel. This implementation was used to build several successful interactive applications as well as batch experiments for research settings. This implementation, in addition to the two new algorithms introduced by this thesis, are a step toward meeting the speed, adaptability, and scalability requirements of practical large-scale vision systems.