1. Feature Detection

a) Harris Detector

Assume structure tensor for each pixel is S. The cornerness response function is R = det(S) - k (Tr(S))², usually k = 0.04 - 0.06.

Without non-maximum supression, the first 2000 corners looks like the following:

Figure 1. Corners detected w/o non-maximum supression

b) After radius 10 non-maximum supression, the results are:

Figure 2.

c) From the following two images, we can see:

1) NCC matching obtain more correct match than SSD matching

2) most matches using SSD focus on trees and grass which has large color difference between left and right image

3) most matches using NCC focus on building and pillar, which has minor color difference.

Matching by SSD Matching by NCC

Figure 3.

2. Two View Image Alignment

a) The inliers obtained by affine RANSAC is

Affine RANSAC find 18 inliers, with projection error of 1.7 pixels Homography RANSAC find 29 inliers, with projection error of 0.2 pixesl

Figure 4.

bc) The result by affine warp and homography warp is shown in the following images.

Figure 5.

a) Align Translation Set

Because there is no skew, perspective etc. component, there is very minor difference between the aligned images.

Affine matching Homography matching

Figure 6.

Align Rotation Set

Affine matching Homography matching

Figure 7.