This part involves using a homography matrix as well as image warping in order to rectify, or unwarp an image. The idea is to take some perspective shape in the input and to morph it into a square in the resulting image. We compute the homography matrix to find this transform using least squares, and apply this transformation to the entire image. The results are shown below:
|
|
|
|
|
|
This part involves using harris corner detection, ANMS, feature detection, feature matching, and ransac in order to autoselect correspondence points for computing a homography.
Using the given harris code, we are able to find a set of corner points in the images. Displayed below is the harris points overlayed on one set of images.
By using ANMS, we are able to surpress the number of points we have down to a certain number. Below shows ANMS run with 500 points.
Using ransac, we eliminate the outliers and keep the best set of inliers to compute our homography.
After finding our best set of points we can compute best correspondences for all of our image sets. From here, we just use our warping implementation and stitch the images together to create a mosaic.