Skip to main content

Graphing Objects

Final Activity Report Summary - GRAPHING OBJECTS (Graphing Objects)

Our main objective is to try to design a computer algorithm for recognising complex objects in images. These objects may consist of several component parts e.g. the human body consists of trunk, arms, legs, head etc. These parts have a set of possible relationships to each other. One way of representing this set of possible relations is by using a mathematical structure known as a "graph", which consists of a network of nodes joined together by edges. The nodes represent different component parts of the object and the edges represent the connections between these parts.

The graph represents a model of the object we are trying to recognise. Our task is to match the model graph to the observed data in the image. We have investigated possible ways to do this based on an analogy with the science of thermodynamics. We consider the discrepancy between the model graph and the image to be a kind of energy and the complexity of the model graph to be a kind of entropy. Our objective is to reduce the energy without allowing an unacceptable increase in entropy. This would result in the optimal match between model and image. We argue that the best way to do this is to represent the graph in a hierarchy of scales of increasing complexity. We start matching the simplest graphs first and then gradually increase the complexity until we reach the optimal trade-off between energy and entropy.

This should produce a more efficient and accurate way of recognising complex objects in images and could be used to recognise the movements of human beings to produce more user-friendly man-machine interfaces.