Whether our objective is object recognition, object reconstruction from measurements, creation of virtual environments, manipulation of objects by a robotic arm, or path planning for a mobile robot, a model of the objects involved is necessary. Having said this, there are different ways in which to represent a given object, depending on the task to perform with it or the agent performing the task. From this point of view, most existing models are myopic in the sense that they are dedicated to a particular task and, often implicitly, a particular agent. Therefore, we have developed a GeNeric Object ModEl (GNOME), which is independent of the task and the agent.
GNOME is a multiple-level model to represent objects through their various attributes (geometry, color, texture, density, etc.). Each attribute is represented independently of the others by a tree of hierarchical decomposition of attribute-objects (objects defined in the attribute's representation space), and by a graph representing relations between these attribute-objects, e.g. connections, relative pose, ordering etc. The set of all trees and graphs for all attributes constitutes the object's structure.
Each attribute-object and relation can be quantified by an application. The application's equation of an attribute-object's quantification specifies the shape (e.g. a cube or a sphere), while values of the equation's parameters specify the dimensions (e.g. the side or the radius). Quantifying the relations consist in linking a function to the relations' graph that indicates the relations and their quantification for each arc in the graph: this gives a labeled graph. The set of all quantifications (of the attribute-objects and relations) constitutes the object's quantification.
Up to now, the attributes have been represented independently, but they are in fact not independent, and inter-attribute relations must be introduced. Since these relations can be quantified, they are represented by a labeled graph. One example of such a relation is correspondence. This relation gives, for example, the color (through color attribute-objects) of the different geometrical parts of the object (through geometry attribute-objects).
Due to the model's modularity and the data structures (graphs and trees) chosen, a representation can easily be modified when it is no longer appropriate (e.g. because of a change of task or agent). For example, attributes can be added or removed or the decomposition hierarchy of an attribute can be reorganized.
We have developed a formalism based on set theory. [LHC96-eng] gives this formalism as well as a more detailed description of the model and a simple example: a car representation for, e.g., a recognition task. I have presented the model in the 64e congrès de l'Association Canadienne Française pour l'Avancement des Sciences [Lab96].
A poster (in French) describes this work.
Go to the description of an underground mine model or back to my home page.