Next: Learning from Human
Up: Experimental Setup
Previous: Test on the
The definition and recognition of cups is a task that has been visited
frequently in machine learning research [Mitchell, Keller, Kedar-Cabelli1986,Winston, Binford, Katz, Lowry1983]. As
Winston (1983) observes, it is hard to tell vision systems what cups
should look like. It is much easier to talk about the purpose and function of a
cup. We convey the description of a cup by providing its functional definition.
In particular, a cup is described as an object that can hold liquid, that is
stable, liftable, and can be used to drink liquids. The physical identification
can be made using this functional definition. In particular, for the synthetic
set of objects created here, these functional properties are broken down into 19
knowledge primitives, 17 of which have range parameters.
We generated a database of 200 synthetic cup examples, for which the measurements
of the knowledge primitives are randomly distributed. Hand-crafted range
parameters (z1,n1,n2,z2) are supplied for all 17 ranges in the cup
functional definition. To generate a cup example, a primitive measurement is
randomly selected for each range. Approximately 80% of the time the primitive
measurement is randomly chosen between n1 and n2. The other 20% of the time
the measurement is randomly chosen outside n1 and n2, but inside z1 and
z2. This cup generator program provides us with the capability to create a
large number of cup examples without the time-consuming process of creating
actual 3-D CAD models for each example.
Larry &
Wed Oct 18 17:48:34 EDT 1995