Project 6: Panorama Stitching

Diana Huang (dkh) - April 19, 2010

Description

For this project, I implemented automatic feature matching and panorama stitching. In order to do this, we used Harris corner detection to find corners in our images. This gave us a dense distribution of corners in our image:

We then used adaptive non-maximal suppression to get a smaller set of corners evenly distributed out over the image.

From these sparse points, we then created file descriptors that allowed us to compare how similar two particular corners were. These file descriptors were normalized and grayscaled to prevent illumination effects from other exposures from influencing the matching.

After creating the file descriptors, we match similar looking descriptors to one another, and then use the RANSAC algorithm to decide which four points to use for the final composite. We can see below that the four points correspond to one another in the final image.

Then, using the four points we found, we then warp the image so that the points remain aligned, and then we composited the images together. For this project, I did not use either Poisson blending or graphcut, as the black edges occasionally made this difficult.

Results Images

For the most part, my images came out well, though there are still some bugs to iron out while working with more than two images.