To view this panorama in full 360, load this applet
This panorama was developed using our MOPS feature detection from project 2. Although the alignment isn't perfect, we thought it was a good demonstration of the effectiveness of our program's ability to deal with sub-par data, as well as a proper illustration of our feature matching program in action.
This subset of a 360 panorama was composed using the alignPairHomography. After a few images, it begins to break down, since it is impossible to continue attaching images together without doing some sort of spherical transformation.
The spherical warping code worked quickly, and proved to be effective with the first iteration. Since matching warped images only requires translations, the RANSAC implementation surrounding that was quickly impemented. Similarly, blending these images together only required translations, and so we did not immediately realize that it did not work for any other type of affine transformation. However, later on in homography testing, we discovered our incorrect implementation and corrected this inconsistency. This led to one unexpected problem, though: how to deal with empty pixels when images get expanded and warped. Our solution was to interpolate, but the initial implementation filled every pixel with an average of its neighbors, leading to a gaussian blur effect. Instead, we ended up extending each pixel to the empty spaces above and to the right, thereby eliminating the blur.