Computational Photography Project 2g: Gradient Domain Fusion

David Wang, Fall 2010

Toy Problem

For the toy problem, I computed the x and y gradients from this image s and reconstructed the image using all the gradients plus one pixel intensity.

Original	Reconstructed	High Contrast	Low Contrast

As intended, the reconstructed image is essentially the same as the original. Just as an experiment, I modified the code slightly to reconstruct the image with greater and smaller gradients to change the resulting contrast, as seen above.

Poisson Blending

For this problem, the goal is to mask out part of a source image and successfully blend it into a background image. To do that, I implemented the following blending constraints:

The results can be seen below:

Source	Target	Pre-blend	Blended




The following example was chosen to illustrate the limitations of Poisson blending. Because of the discrepancy in color and intensity between the background behind woman and the background image, and because the area of the background image that the woman is being pasted into does not have a relatively uniform intensity, Poisson blending fails to blend it in a reasonable way.

Mixed Blending

This problem is similar to Poisson blending, but instead of using the source gradient as the guide, I use the gradient in the source or target with the greater magnitude. These are the blending constraints to be implemented:

The results can be seen below:

Source	Target	Poisson Blending	Mixed Blending

As is evident, mixed blending works far better for things like blending writing or text with a uniform background into a highly textured image.