20.01.2013 Views

Master Thesis - Department of Computer Science

Master Thesis - Department of Computer Science

Master Thesis - Department of Computer Science

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

Table 4.1: Effect <strong>of</strong> increasing number <strong>of</strong> training samples on the performance <strong>of</strong> null<br />

space and range space for Yale database.<br />

No. <strong>of</strong> Training Null Space Range Space<br />

Samples<br />

2 93.33 64.44<br />

3 93.33 76.67<br />

4 90.48 82.86<br />

5 88.89 81.11<br />

6 85.33 76.67<br />

7 81.67 76.67<br />

8 75.56 75.56<br />

9 50.00 56.67<br />

attains a maximum value 95.00% for eight training samples but again decreases down<br />

to 92.00% for nine training samples due to the small size <strong>of</strong> null space. The perfor-<br />

mance <strong>of</strong> range space exhibits an interesting behavior. The accuracy <strong>of</strong> range space<br />

increases with increasing number <strong>of</strong> training samples for the following two reasons:<br />

(i) the subspace learns more and more about the pose variations across the database<br />

and (ii) discriminative informations go to range space due to the increase in the<br />

number <strong>of</strong> training samples. Thus maximum accuracy for range space is obtained as<br />

97.50% for eight training samples. But the further inclusion <strong>of</strong> more training samples<br />

reduces the performance due to the following reasons: (i) the new training samples<br />

does not provide any extra discriminatory information with respect to the informa-<br />

tion already learned by the classifier from previous training samples, (ii) moreover,<br />

they add confusing information to the discriminatory features.<br />

For PIE database the performance <strong>of</strong> null space and range space is evaluated<br />

and shown (see Table 4.3) for even number <strong>of</strong> training samples ranging from two (2)<br />

to eighteen (18). The performances <strong>of</strong> null space and range space with increasing<br />

number <strong>of</strong> training samples can be explained by a similar logic as described in case<br />

<strong>of</strong> Yale and ORL databases. PIE has only illumination variation.<br />

93

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!