Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
Master Thesis - Department of Computer Science
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
Table 4.3: Effect <strong>of</strong> increasing number <strong>of</strong> training samples on the performance <strong>of</strong> null<br />
space and range space for PIE database.<br />
No. <strong>of</strong> Training Null Space Range Space<br />
Samples<br />
2 80.21 40.17<br />
4 82.58 72.36<br />
6 88.33 75.28<br />
8 90.44 89.75<br />
10 93.49 97.55<br />
12 55.33 94.56<br />
14 53.87 94.44<br />
16 50.58 95.90<br />
18 39.86 96.04<br />
robust and efficient for classification for higher number <strong>of</strong> training samples. The<br />
common features extracted from a number <strong>of</strong> faces having expression variation<br />
does not help in classifying the samples from the same class showing illumina-<br />
tion variation. So, the performance <strong>of</strong> null space on Yale database is high with<br />
very less number <strong>of</strong> training samples. On the other hand, the variability among<br />
samples contributes in a different way in case <strong>of</strong> range space. Higher number<br />
<strong>of</strong> training samples contributes to a large amount <strong>of</strong> intra-class variations in<br />
range space. Thus, the information learned by range space from larger number<br />
<strong>of</strong> training samples does not help in discrimination and adds to the degree <strong>of</strong><br />
confusion.<br />
• On the contrary, ORL and PIE databases contain only a specific type <strong>of</strong> vari-<br />
ability (pose for ORL and illumination for PIE). So, higher number <strong>of</strong> training<br />
samples provides robust common features for null space and <strong>of</strong>fers discrimi-<br />
native information (instead <strong>of</strong> adding confusing information) to range space.<br />
Hence, the performance increases with higher number <strong>of</strong> training samples in<br />
the case <strong>of</strong> these two databases, unlike that in Yale database.<br />
95