Combining Text Semantics and Image Geometry to Identify Relations
Dennis Medved · Lund University Publications (Lund University) · 2012
Automatically identifying actions and relations between objects in images can be useful for many applications, for example image and video labeling.In this thesis, the goal is to extract a predefined set of objects from the images and identify the relations linking these objects.Wikipedia is the source of images and texts that are analyzed.I created a program that takes both geometrical information from images, and semantic information from texts and outputs classification of relations for each object in the images.The baseline of only using geometrical features is improved on by using bag-of-words features based on the articles, and further enhanced by utilizing predicate information.Pierre Nugues: thanks for being my supervisor and