I3GO+ at RICATIM 2017: A semi-supervised approach to determine the relevance between images and text-annotations

Abstract

In this manuscript, we describe our solution for the RedICA Text-Image Matching (RICATIM) challenge. This challenge aims to tackle the image-text matching problem as one of binary classification, that is, given an image-text pair. Therefore, a valid solution must determine if the relation between the image and text is valid. The RICATIM dataset contains a large number of examples that were used to create an algorithm that effectively learns the underlying relations. Vision and language classifiers must deal with high dimensional data; therefore, traditional classification methods increase their learning time and also tend to perform poorly. To tackle the RICATIM challenge, we introduce a novelty approach that improves the classification based on k-nearest neighbor (KNN) classifier. Our proposal relies on the solution of the k centers problem using the Farthest First Traversal algorithm, along with a kernel function. We use those techniques to reduce the dimension effectively while improving the performance of the KNN classifiers. We provide an experimental comparison of our approach showing a significant improvement of state of the art.