Push and Pull: Iterative grouping of media

While many techniques use the traditional approach of time consuming groundtruthing large amounts of data [1, 2], this is increasingly infeasible as dataset size and complexity increase. Instead we propose a solution that allows the user to select media that semantically belongs to the same class and use machine learning to “pull” this and other related content together. We use real data harvested from the internet and propose an approach capable of incrementally clustering similar material using the manual identification of a few true positive and false positive examples. In order to provide both scalability and incremental learning, the approach needs to be efficient. We combine two popular data mining tools developed for the text analysis domain to efficiently compute distances between high dimensional representations and dynamically augment the representation with new compound visual words to form an image signature. These tools are applied to selected true and false positive examples of the media and rules learnt, that are applied to the full corpus of material. The media is then formed into groups of same class media using a greedy clustering approach. An image signature is constructed for each input sample; this is similar to a bag-of-words (BoW) histogram representation, and provides a compact, discrete representation of the input sample, as shown in Figure 1. In order to form the similarity between the image signatures, a data min-