In this experiment, 660 training sequences were used to build a recursive partition tree. For each nonterminal region, we selected an adaptive radius r as defined in Section 4.4 to split the region into subregions. Given r for a nonterminal region, we have k>1 training samples and the distance between each pair of these k samples is greater than r. These k samples were the centers to generate one level partition. The distance here is the Euclidean distance in the MDF space corresponding to the region. Fig. 15 shows the top 10 MDF's at the root level. If we compare the MDF's in Fig. 15 with the MEF's in Fig. 14, it is clear that the top MDF's capture feature locations (edges) because it accounts for the major between-sign variations.
Table 4: Summary of the experimental data for RPTA
Once we have created the recursive partition tree, we used it to recognize the sign. As we did in the experiments for the nearest neighbor approximator in the MEF space. The segmentation result was used as the input for sign recognition. The results are summarized in Table 4. The correct recognition rate of 161 testing sequences is 93.2% which is better than the recognition rate (87.0%) of the nearest neighbor approximator in the MEF space. The average recognition time per sequence is 0.63 second on the SGI INDIGO 2. The time is longer than the time (0.27 seconds) of nearest neighbor approximator when the quasi-Voronoi diagram is used in the query. This is because each nonterminal node in the recursive partition tree has its own version of DKL projection matrices, each requires a projection, whereas in the case of the nearest neighbor approximator, we only need one projection to the MEF space.