Leave-one-out test results for the real-object database with evaluation measures derived from human ratings of the objects are listed in Table 1. Recall that the error rates are not directly comparable among the three categories.
Table 1: Leave-one-out test results for real-object database with evaluation
measures derived from human ratings of the objects.
The actual evaluation measures for the conventional chairs objects are within approximately 7% of the human evaluation measures. The average error here is about 6% greater average error than for the GRUFF data with a similar number of training samples. The histogram in Figure 16 A shows that the data set of real conventional chair objects contains mostly ``good" examples. Thus, the higher average error can probably be attributed to the ``noise" associated with the real-object evaluation measures. Considering an average standard deviation of 12% for the human evaluations of the conventional chair objects, a 7% average error per sample for the OMLET results does not seem unreasonable. The actual evaluation measures for the real-object straightback chairs and armchairs differ on average by less than 1% from the desired measures. As before, all conventional chair examples were used to train the ranges associated with the conventional chair category before the ranges for the straightback chair category were trained. The histograms of desired evaluation measures for the back support of the real straightback chair objects and the arm support of the real armchair objects are shown in Figure 16 B and C, respectively.
Figure 16: Histograms of desired evaluation measures of the
real-object training sets.