Project 4: Scene Completion (Writeup)

Jason Pacheco (pachecoj)

March 12, 2010

Algorithm Description

Source Image
Align Matches
Best Alignment
Graph Cut + Blend

The images to the right illustrate the steps of the algorithm. First we are given a source image and a mask (mask not shown). We then align all of the matches provided and choose the one with the lowest SSD score of RGB values between the source and match. Then we constrain the pixels inside the mask to come from the match and we form a mask on the source that constrains all pixels outside of a 100px boundary around the input mask to come from the source. Therefore we have a 100px boundary around the input mask of unconstrained pixels. We perform a graph cut and Poisson blend to obtain the final result.

Graduate Credit

Multiscale Alignment

The alignment is done similar to the first project with modifications to allow different-sized images. However, as in the paper, we look at multiple scales of the matched image and align each. We choose the alignment with the lowest SSD score for RGB values between the source image and the match. As in the paper we choose the scale factors (1.0, 0.9, 0.8). This modification shows significant results in a few images.

Modified Graph Cut Cost Function

We use the graph cut cost function described in the paper,

This cost function shows a significant improvement in some images with strong edges by encouraging cuts that do not cross these edges. This is especially noticeable in images with water and a shoreline since the cut will typically be unconstrained in the water, but tend to cut along the shoreline.