Free Trial

Safari Books Online is a digital library providing on-demand subscription access to thousands of learning resources.


Share this Page URL
Help

LEARNING A DISTANCE MEASURE: THE NEW APP... > LEARNING A DISTANCE MEASURE: THE NEW... - Pg. 179

Annotating Historical Archives of Images particular heraldic shields dataset. This is because these objects have identical outlines (within the limits of the artist's ability) whereas their colors are very diverse. Therefore it is not surprising that we have greater accuracy when we employ only the color measure. We used the algorithm Join_by_Combined_ Measure shown in Table 2 to combine the color and shape features on heraldic shields dataset, where X is the 100 drawn heraldic shields images and Y is the 2,350 synthetic heraldic shields, and w max = 1. In the absence of ground truth we divided the results into 3 categories, perfect matches, not perfect but plausible matches and poor matches. Figure 21 shows examples of each type. In total, we had 16 perfect matches, 19 plausible matches and 65 poor matches. The vast majority of the poor matches are simply patterns which did not exist in our database, or where the hand-drawn historical image was highly degraded. Exploiting Chirality of Projectile Points In the two previous examples we had a situation where we expected using just color to be best (heraldic shields), and a mixture of color and shape to be best (butterflies). For completeness we now consider a dataset where we strongly suspect that only shape matters. In particular we examined a set of drawings of projectile points (arrowheads) from a field archeologist's notebook (Gregory, 1958). While the sketches capture the shape (and to some degree the texture) of the objects in question, they are clearly devoid of any color information. In this experiment, 30 images of hand-drawn arrowheads, taken from historical documents (Gregory, 1958), are used as the set A for the al- gorithm Learn_Weighting_Parameter. As before, we produce the set B by reversing the 30 drawn arrowheads of set A from left to right. Figure 22 shows when w goes up above 0.5, the classification accuracy drops dramatically. The figure implies that the shape information is of greater importance than the color information in this application. Note that we have the problem here of break- ing ties, since the accuracy is maximized over the range of w = [0, 0.5]. Ties may be common using our technique, given that the range of values for the classification accuracy is a relatively small integer (i.e., |A|). We can break ties by choosing the value in the maximizing range of w that minimizes the sum Figure 22. Classification accuracy on the arrowheads dataset when the set B is obtained by exploiting chirality 179