5. Match Space

A simple matching algorithm was conceived and implemented to recognize Doc, given a set of blobs represented by boxes in a list. The algorithm is as follows:

1.

Remove all but the blob with the largest area from each region, so that five large blobs remain.

2.

Compute a histogram of distances from the center of every blob to the center of every other blob.

3.

Compute the centroid of the first three blobs which are about the same distance apart - this is Doc's location.

In practice, this matching technique doesn't do all that good a job at finding Doc. A very large blob which may not correspond to a part of Doc often results in less than optimal classification. In the test scene of Figure 17, the matches were among the beard, suit and hat, leaving a mean on the lower portion of Doc's face. In Figure 18, the match is rather successfully identified near the exact center of Doc. Figure 19 is a suprisingly good match, despite that the wall was utilized as a match for Doc's beard. Finally, Figure 20 shows a poor match for Doc thrown off by the white wall and the yellow square on the fire extinguisher in the background.

  

Figure 17: Results

  

Figure 18: Results

  

Figure 19: Results

  

Figure 20: Results