Matthew Banner

15-463 Fall 2007

Project 2: Content-Aware Image Resizing

Energy Function Comparison

I tried two different energy functions: a simple gradient function and a correlation using the Laplacian of Gaussian. You can see a comparison of these two below. Notice that the images produced using the gradient function tend to have their geometry preserved but sometimes contain sharp edges where seams have been removed. The LoG function, on the other hand, seems to produce smoother borders between different sections of the image, but often distorts the geometry thereof. I find images produced using the LoG function to be preferable, so all images shown here other than the comparison below were created in that manner.

The largest images are not shown so as not to distort the layout of the page, but it's really the narrower ones that are interesting. This image was chosen for the comparison because it contains both elements that are handled effectively by the algorithm as well as some which are not.

Laplacian of Gaussian Gradient
Original (600x450) Original (600x450)
LoG (550x450) Gradient (550x450)
LoG (500x450) Gradient (500x450)

Successful Images

In this image, the algorithm successfully removes the sky as well as some of the beach while leaving the abondoned ships mostly intact (source).
Although the train is somewhat compressed in this image, notice how much more the back end, where it is hard to distinguish from the background, is than the sharper, more distinct front end (source).
Notice how the sky is first removed, followed by the sand. This leaves the trains in the middle nearly untouched except that they now appear to be resting along a ridge (source).
It would not have been a true computational photography project without the Russian Guy in the Blue Shirt. Notice how the featureless wall is removed, leaving the subject completely intact (source).
Here, the height of the image is reduced without removing the highlights of the bridge's reflection (source).
We are able to greatly reduce the height of the image without losing the details of the sky (source).



Unsuccessful Images

Unfortunately, the least interesting portions of this photo (according to my energy function) were the towers of the church. This led to the unfortunate tendency of the structure to look as if it were melting (source).
These look pretty good as long as you don't compare the trunk of the tree with its reflection (source).

The End