15862 Computational Photography
Project 2

Yiling Tay (ytay)

Poisson Blending

Overview

The goal of this assignment was to use gradient domain manipulation to create interesting blends and effects with photographs. Specifically, we blend different features from different images together, to create a realistic looking result.

This is done by solving a least squares equation to minimize the differences in gradients between the target (foreground) and source (background). Here, both source gradient and mixed gradients are used. In the case of mixed gradients, the max of the source and target gradients are used. By forcing the gradients of the source and target to match, the output results in a smooth blend between the two images.

Results:

Toy Problem (Sanity test to recover original image)
InputOutput


Mouthy Man

UFO Sighting

Call for Help (mixed gradients)

Watchful Eye (mixed gradients)

Failure Case

The vast difference in lighting in the forground and background, as well as difference in texture of the skin, causes this blend to look unrealistic.


Bells and Whistles: Color2Gray

When converting a colored image to grayscale, some important contrast information disappear, as shown below. By using the HSV space of the colored image, we can convert the image to grayscale by preforming a mixed gradient blend, using the contrast information from the saturation and value spaces as the target, and the actual grayscale image as the source.

Original Matlab's rgb2gray Results of Color2Gray



Yiling Tay - Carnegie Mellon University - Fall 2011