Sensitivity to image added noise

The SIFT descriptor is sensitive to the image added noise, as we can see in figure 5.6 on page . The classification rate was quite stable up to certain $v$, $v=0.1$. At this point the clasification rate rapidly decreased and simpler SIFTs achieved better classification rates for $v=0.3$ in comparison to SIFT-GradWei SIFT. The initial classification rate was around 90% for $v=0.02$, for $v=0.1$ it was 88%, but for $v=0.3$ it was only 66%.

Figure 5.4: Images after noise addition. (a) $v=0.02$, (b) $v=0.3$

Figure 5.5: k-NN classifier sensitivity to image added noise

Figure 5.6: In this picture we can see that SIFT discrimination power decreases with increasing amount of noise. For $v$=0.3 the distribution is almost random (d); (a) original image, (b) added noise, $v=0.04$, (c) added noise, $v=0.1$, (d) added noise, $v=0.3$

Table 5.3: Sensitivity to the image added noise(see figure 5.5 on page )

noise ($v$)	SIFT-GradWei	SIFT-Grad	SIFT-Ori
0.02	91.3	89.13	82.61
0.04	92.39	89.13	82.61
0.06	89.13	83.7	80.43
0.08	88.04	83.7	78.26
0.1	88.04	83.7	75
0.2	69.57	70.65	69.57
0.3	66.3	69.57	56.52