Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
4.5.1 Effect <strong>of</strong> Number <strong>of</strong> Training Samples on the Performance <strong>of</strong> Null<br />
Space and Range Space<br />
To demonstrate the effect <strong>of</strong> increasing number <strong>of</strong> training sample on the performance<br />
<strong>of</strong> null space and range space, we compute the recognition accuracies <strong>of</strong> null space<br />
and range space for Yale, ORL and PIE databases with different number <strong>of</strong> training<br />
samples. For Yale and ORL databases, the accuracies are calculated for the number<br />
<strong>of</strong> training samples per class ranging from two to nine. As the number <strong>of</strong> training<br />
samples per class for PIE database is higher (42), the number <strong>of</strong> training samples is<br />
varied in the range 2 to 18 (e.g. 2, 4, 6,..., 18). The image size has been kept as 25*25<br />
for all databases. The null space technique used in our experimentation is adopted<br />
from the method called Discriminative Common Vectors (DCV) [20].<br />
The performance <strong>of</strong> null space and range space for Yale database is shown in<br />
Table 4.1. The performance <strong>of</strong> null space is best when the number <strong>of</strong> training samples<br />
per class is only two and three. Performance <strong>of</strong> null space decreases with increasing<br />
number <strong>of</strong> samples and provides the minimum accuracy <strong>of</strong> 50% when number <strong>of</strong><br />
training samples is maximum (nine). This result validates the claim <strong>of</strong> negative<br />
effect <strong>of</strong> increasing number <strong>of</strong> training samples on the performance <strong>of</strong> null space.<br />
Initial increase in the number <strong>of</strong> training samples enhances performance <strong>of</strong> null space,<br />
because the captured common features for each class hold more robust commonalities<br />
present across samples containing variations in expression, illumination and pose. In<br />
case <strong>of</strong> range space, the performance increases with increasing number <strong>of</strong> training<br />
samples upto a certain point and again deteriorates. This performance trend can<br />
be easily explained by stating that initial increase in number <strong>of</strong> samples drives the<br />
discriminative information to the range space. However too many number <strong>of</strong> training<br />
samples leads to a huge amount <strong>of</strong> intra-class variations which in turn effects the<br />
overall accuracy.<br />
The performances <strong>of</strong> null space and range space on ORL database with training<br />
samples varying from two to nine are shown in Table 4.2. As the ORL database has<br />
only pose variation, the increasing number <strong>of</strong> training samples helps the null space to<br />
capture the common features which are robust and efficient for classification. So, the<br />
performance <strong>of</strong> null space increases with increasing number <strong>of</strong> training samples and<br />
92