CS 766 Final
Samareh Shahmohammadi
When Images Don't Line Up
In order for this process to work well, the images from different exposures need to be, pixel for pixel, of the same scene points. When they don't, the result will not be good.
Image Registration
The process of reading images and lining them up is called image registration.
​
Matlab image processing toolbox has a registration function, and I tried to
use it. However, it needs a lot of tweaking to line up well and takes a long
time to process. So I tackled the problem of lining up images using
Homography:
​
-
Generate SIFT matches to find corresponding points between images
-
Eliminate erroneous matches using RANSAC
-
Compute homography based on the matching points
-
Warp image using the computed homography
​
I chose to start with the “middle image” (as described on the Automated Point Selection page). Then, I and warp consecutively from there with the assumption that consecutive images with the closest exposure times will have best match. The results are below:
Courtesy of Mohit Gupta
The results look fairly good. I discovered that the homography registration is actually much faster to compute and more accurate than the image processing toolbox tools.