Sensitivity to image blur

It can be observed from the chart (figure 5.2 on page ) that SIFT was almost invariant to blur. The possible explanation is that SIFT shows almost no change in feature vector in blurred image. (figure 5.3 on page ). The classification rate was around 92% for the SIFT-GradWei and around 82% for SIFT-Ori.

Figure 5.2: k-NN classifier sensitivity to image blur (truck images)

Figure 5.3: SIFT is almost invariant to blur (a) original image, (b) Gaussian with 6x6 region size applied on image

Table 5.2: Sensitivity to the image blur(see figure 5.2 on page )

Gaussian kernel size ($\sigma$)	SIFT-GradWei	SIFT-Grad	SIFT-Ori
0.5	92.39	90.22	81.52
1	92.39	90.22	81.52
1.5	92.39	90.22	81.52
2	92.39	89.13	81.52
2.5	92.39	89.13	81.52
3	91.3	90.22	80.43
5	90.3	-	-