15-463 Project 4 Parts 1 and 2: Automatic Image Warping and Mosaicing

David Klionsky

Rectified Images

Here are some pictures I've taken that have been rectified. The first is the UC colonnade, the second is some graffiti I found in New York.

UC Colonnade

Graffiti

Part 1: Manual Mosaics

Correspondence points for these mosaics were set by hand. The blending is not perfect, and some edge artifacts are visible. Automatic matching of the backyard pictures is in part 2 below.

Backyard

Stairs

The Cut

Part 2: Automatic Mosaics

To automatically make a mosaic, interest points are first detected in the source images. A Harris corner detector was used to detect these interest points. On the left are the top 500 strongest interest points. On the right are the top 500 points as found by adaptive non-maximal suppression.

8x8 pixel feature descriptors are made of each interest point, using a 40x40 pixel window around the point. Here are the feature descriptors from the right image above.

Feature descriptors are made for interest points in both images, and the descriptors are then matched. Descriptors that do not have a good match are discarded, but some incorrent matches are made.

RANSAC is used to eliminate the incorrect matches.

The points that are left are correctly matched and are used to compute a homography between the two images. The left image is warped on to the center and the result is shown below.

David Klionsky, 3/19/2010