Fisher's linear discriminator (FLD) performed on car images

Experiments for Car Images using FLD were done on SIFT with topology 15x9x15 and overlapping tiles since this was the best combination obtained so far for car images. For truck images the optimal SIFT was 15x9x25 and with overlapping tiles, however the computation was so complex that we never obtained the projection matrix within 12 hours. We decided to use SIFT with topology 15x9x13 without overlap and compare this to the k-NN with the same SIFT topology.

Here we can see the FLD projection from SIFT feature space to training set projection space. We can see in figure 4.15 on page the distribution of classes and in figure 4.16 on page the distribution of truck classes when using FLD.

Figure 4.15: The FLD transformation performed on features extracted from car images

Table 4.10: Classes to numbers assignment(see figure 4.15 on page )

Car Class	Car Label in the picture
octavie	1
opel	2
opel2	3
skoda120	4
felicie	5
skoda105	6
favorit2	7
fiat-uno	8
favorit	9
fabie	10

Figure 4.16: The FLD transformation performed on features extracted from car images