Affordance Features

AfNet uses three types of affordance features. These are structural (primary), material (secondary) and semantic (tertiary) affordances. Structural and material affordances together form functional affordances. While structural affordances hold the key to object class recognition in visual perception systems, material and semantic affordances accelerate recognition while improving confidence levels for one of many alternate fine-grained object identifier hypotheses. In certain cases, based on uniqueness and discernability, material affordances can also be used independently for object recognition. Structural affordance corresponds to inferred knowledge about the 2D/3D shape of the object. For example, detection of a cylindrical or circular shape corresponding to a part of the object or as a whole, indicates a 'Roll-ability' affordance. Material affordance corresponds to deduced knowledge about the material properties of an object based on the local visual features. A good example is the shiny or grey color of a metallic object which results in an inference of high strength and 'dur-ability' of the object. Semantic affordance is dependent on co-occurrence or typical pose of objects. Semantic affordances define the Subject-Object relationships for functional affordances. The current ontology of functional affordance features (currently 25 structural and 10 material) used by AfNet is presented in the following chart.

AfNet

An Example

An example for affordance features is the 'Contain-ability' structural affordance feature. This feature is exhibited by objects or object parts that provide the functional ability to contain in stable or ground state, a solid or liquid within its geometry. AfNet describes unique geometric mappings for each affordance feature. Contain-ability is defined using geometric mapping of 'high convexity'. Mugs, cups, beakers, bowls, bags and pots exhibit this functional affordance and hence form a 'Conceptual Equivalence Class'. But fortunately, to the dismay of topologists, donuts do not fall under the same category, as demanded by homeomorphism, enabling legitimate functional visual categorization.

AfNet