Deep Belief Networks (DBNs) are capable of learning the structure of highly complex data without supervision. When applied to images, however, DBNs have difficulty representing object or texture boundaries since these represent a transition from one set of image statistics to another that is not well captured by a fully-connected DBN layer. We present the Deep Segmentation Network which allows such boundaries to be modelled by factoring out the appearance of an image region from its shape, and using separate pathways for each. Such segmentation is applied at every level in the hierarchy. We present results on decomposing low level appearance and shape in natural images and discuss how deeper models could be used to segment and recognise object parts, objects and scenes by considering segmentation at different depths. This is joint work with Nicolas Le Roux, Jamie Shotton and Nicolas Heess.
BACK
