Problem Set 5
CMSC 426
Assigned Monday, March 15, Due Thursday, April 1
The goal of this problem set is to implement the texture synthesis method of Efros and
Leung. This is described in the paper: ``Texture Synthesis by Non-parametric
Sampling’’, by Efros and Leung, in the International Conference on Computer Vision,
1999. There is a link to this paper on the class web site. Below we also include
pseudocode for the project, which is given on their web site along with some samples of
the system’s performance. I have simplified their pseudocode a little bit, to remove some
details such as Gaussian weighting which don’t seem to be necessary to achieve good
performance.
This algorithm takes a sample of some texture and generates a new image containing a
similar texture. The strategy of the algorithm is to generate each new pixel in the image
using a neighborhood of already generated pixels. One looks in the sample for similar
neighborhoods, selects one of these similar neighborhoods at random, and copies the
corresponding pixel into the new image. You are only required to implement this
program for black and white images.
1. 30 points: To begin, you should write a program called my_SSD which performs
the central step of the algorithm. This will compute the sum of squared difference
between a little portion of the new image you are making and every portion of the
sample. It should have the form:
function D = my_SSD(S, T, M)
S will contain a sample of the texture we want to generate. T contains a small
nxn neighborhood of pixels. Not all the pixels in this neighborhood have been
filled in with valid values however. So M (the mask) is an nxn matrix that
contains a 1 for each position in which T contains a valid pixel, and a 0
whenever the corresponding pixel in T should be ignored. Computing the SSD
is like correlation in that we shift the template over every position in the sample,
and compute a separate result for each position. Thus, the output D is the same
size as S. To compute D(i,j) we shift T so that its center is right on top of S(i,j).
Then we take the difference between each valid pixel in T and the corresponding
pixel in S, square the result, and then add all these together. This computation is
described on page 147 of Trucco and Verri, and by Efros and Leung with the
following fragment of pseudocode:
