My implementation allows a mosaic to be created automatically from properly-named files. In this case, I took advantage of that capability to change the "pinned" photo to effect a change in perspective. Respectively, these images simulate looking left, center, and right in the Morewood Gardens courtyard.

I took these pictures because I wanted to try a vertical mosaic, since I figured most submissions would be wide panoramas of landscapes, etc. This photo really makes the perspective effect obvious, though, unfortunately, my "idiot" camera (though ideally suited for an idiot photographer such as myself) did not allow keeping constant settings through these shots. Also, I probably took more photos than were necessary to make this a passable mosaic.

An odd apparatus we set up freshman year to dim the lights in the Hammerschlag House lounge. The second image has been warped such that the third dimmer box from the left is a rectangle.

The first step is identifying Harris corners in the image. There are 4860 here.

Next, adaptive non-maximal sampling (ANMS) is applied to find representative Harris corners that are uniformly distributed across the image. Here, the number of points has been reduced tenfold, to 500.

The same ANMS process applied to another image, in anticipation of trying to match the two.

Samples are taken at a 5-pixel stride from a 35x35 window around each feature point. The samples are taken from a blurred version of the image to assist in matching. The samples are normalized such that their means are zero and their standard deviations are one. Points are then matched between images using a simple distance metric, where matches are only held as valid if the best match is "sufficiently better" than the second-best match. Notice that the resulting pairings correlate very well, but that there are still a few outliers. Here, there are 181 matching pairs.

Applying a RANSAC (RANdom SAmple Consensus) to the image allows the pairings to be aggregated into "agreeing" groups, where the largest agreeing group is taken to be the correct correlation. From this pairing, a homography is determined for each pair of images, and warping is performed as before. There are now 139 matching pairs.

Automatic alignment using hundreds of points, rather than a dozen, makes for much cleaner stitching. Also, a bug was fixed from part I that caused alpha blending never to occur. A new feathering algorithm was added to lighten the transition between edges.

The same vertical panorama as above, only with fewer images, automatic alignment, and feathering.

A park seen from Forbes Avenue near the base of the Cathedral of Learning. Heinz Chapel is on the left.