Modelling of a Bidirectional Network

Abstract

Visual place and object categorization has been an important aspect of research for a number of years now. One reason for its popularity is the wide number of applications in Mobile Robotics and HRI. Some of these include, behavior based navigation, mapping, task based planning, semantic SLAM and active object search. Recently, there has been a trend in place categorization based on the objects associated with that place. The reason that object based place classification proves useful is that a standard camera image represents partially observable environment at any one point in time and where simple place based classification will fail in such an environment, object based place classification is more successful. Inspired from the successful results of these algorithms, this research formulates a new approach where the correlation between two input types aids the better classification or categorization of both the inputs. The purpose of this research is to contribute to simultaneous place and object classification. This research augments the ongoing research in three areas; i) A novel method is introduced for bidirectional classification ii) A randomized object detector and localizer is introduced iii) The approach shows superior performance to separate predictions of object and place classification.