Evaluation of a Zero-shot Cross-modal Image Retrieval Technique Using Ranking Support Vector Machines

Abstract

The aim of this thesis is to evaluate the potential of a technique for cross-modallabel based retrieval of images, from previously unseen classes, using a rankingsupport vector machine and investigate attributes of the system that effect performance.Ranked retrieval performance of the method was measured using mean AveragePrecision and precision-recall. Monte Carlo simulations were used to compareagainst random chance retrieval and provide a baseline for performance. Commonlyused and freely available datasets were used for training and performanceevaluations. Several training attributes and the learned weights of the system wereanalyzed to better understand their effects on performance and characterize themethod.While comparison to other zero-shot methods are difficult to draw the results werecomparable with the simulations and one similar technique evaluated on seenclasses. The testing and simulations showed that the method is capable of betterthan random retrieval of unknown classes from a large database and has superiorperformance on retrieving seen classes than a recent comparable technique.Investigations into different attributes of the system, particularly training sizes, indicatesthat the method may benefit greatly from using a large amount of data fortraining.