SlideDeckFinder: identifying related slide decks based on visual appearance and composition patterns

This paper introduces SlideDeckFinder, a tool integrated into a user's email client enabling the search for similarities between slide decks. The similarity calculations are based on visual correspondence (both from text and images/graphics) as well as slide (re-)composition patterns. The individual slides of different slide decks are first compared by matching their respective visual features extracted from any content such as text and images. The resulting similarity scores between pairs of slides are then the input for calculating the similarity between whole slide decks. Hidden Markov models (HMMs) are used to represent the transformation (in terms of re-arrangements or insertions of new slides) from one slide deck to another, where the state emissions probabilities of the HMM correspond to slide similarity and the transition probabilities represent the likely slide sequence within slide decks. The Viterbi algorithm is finally used to calculate the most likely state sequence (i.e. recomposition pattern) between the slide decks and thus the similarity score. SlideDeckFinder has been evaluated both on its accuracy to compare visual appearance of slides with respect to human perception and its performance to retrieve related slide deck variants.