Hybrid Images

Pyramids

When constructing the pyramids, I uniformly used a pyramid of size 8. For each layer, I blurred the image using the fspecial gaussian blur with a 9x9 kernel and a symmetric edge pattern. The sigma fluctuated between 1 and 1.5 depending on the image.

When downsampling, I downsampled from the original image, not from the blurred one. This is because I used imresize which blurs before downsampling. For each layer I resized by 1/2. The corresponding layer of the laplacian pyramid is the current image (downsampled but not blurred) minus the most recent gaussian (downsampled and blurred).

Hybrid Image

When computing the new image from the pyramids, I start with the gaussian image at cutoff1 of the pyramid. I resize to the size of the original image. From there I sum 0-cutoff1 of the laplacians from the second image, and cutoff2-n (in my case, 8) of the first image. Cutoff1 and cutoff2 change with each image.

Sadly there are no bells and whistles. I got too excited making hybrid images from the staff page of CS17. If you are interested in seeing Scott Mitchell mashed with Gob, let me know (his came out the best).

Results (in increasing level of hilarity)

This is what their babies would look like

this is what our babies would look like

Discussion

The ones that worked well were images with large, uniform-colored backgrounds. Then the more detailed images were better in the front. Fiddling with the cutoff parameters and the gaussian kernel size and the sigma value was fun.