Figure 9: The 52 object chair database.
From the evaluations of GRUFF [Stark Bowyer1991], a large database of 3-D shapes specified as polyhedral boundary representations has been built up. Figure 9 shows 52 chair shapes. A number of the 52 shapes can belong to more than one category or can function in more than one stable orientation. This results in a total of 110 training examples. There are 78 labeled instances for the category conventional chair. Some 28 of these instances additionally satisfy the function of straightback chair, and 4 instances satisfy the function of armchair. For each shape, we have the evaluation measure for the shape's membership in different object categories, as computed by GRUFF with the hand-crafted functions for the primitive evaluation measures. This set of shapes and their evaluation measures make up the first set of training examples.
The first set of experiments will help determine how well OMLET learns a set of membership functions that minimize the overall error, and also how closely the learned membership functions approximate the original functions hand-crafted by an expert for GRUFF. A question of great practical importance to vision researchers is whether a machine learning technique can derive a set of system parameters equivalent to the hand-crafted results of the system designer. If so, the manual effort in system construction could be greatly eased. When the learning task is formulated as duplicating the GRUFF measures, the training data for these experiments is effectively ``noiseless". (Noiseless in the sense that the desired evaluation measures that are used as input to OMLET are all derived in the same manner from the same set of hand-crafted fuzzy membership functions.)