In order to demonstrate the automatic selection of histograms and parameters for inclusion in the object recognition database, we have specifically constructed sets of objects which include representational issues that need to be automatically resolved. For example, when selecting representative histograms, we would like the solution to include examples constructed from noise free data, even though more noisy examples were present. Also, we would wish the representation chosen for each histogram to be driven according to the degree of invariance characteristics in the variant sets provided. Finally, we would want the selected histograms to be based on the most complete version of an object. An automatic object recognition system with these characteristics would then be capable of learning to recognise examples of objects in cluttered scenes (as humans can).
Figure 6:
Two Groups of Data Used
The algorithm was tested on two groups of data containing 1207 lines and 2699 lines respectively. The first group of data consisted of the four data sets shown in Figure 7 . In each of these sets the leftmost object is that shown in Figure 6(a) . The pterodactyl set of Figure 7 consists of variants having the original line length shortened or lengthened by a different factor. The second set of Figure 7 consists of six objects the first is the original stegosaurus, the following two are the first with increasing levels of Gaussian noise added. The last three stegosaurus objects are the first three mirrored around the midpoint of the objects length. The third data set consists of six dimetrodons having increasing levels of Gaussian noise added to the lines. Finally the triceratops set contains objects all having different segmentation accuracy, giving various degrees of a coarse polygonisation effect.
Figure:
Dinosaur Data Sets
(From top-left going clockwise: pterodactyl, stegosaurus, triceratops
and dimetrodon)
The second group of test objects consisted of a duplo person, cow and horse and are shown in Figure 8 . In each of the sets of Figure 8 the leftmost object is that shown in Figure 6(b) . The persons of Figure 8 were produced by shortening and lengthening the lines of the original duplo person. The second cow of Figure 8 is the first cow with gaussian noise added and the third and fourth cows are the first two mirrored around the midpoint of the objects length. Finally the horse set consists of horses having different accuracies of segmentation.
Figure 8:
Duplo Data Sets: person, cow and horse
Frank Aherne