The program submitted to you all is rather simple in theory.  Essentially,
I use a open source feature detector on all of the original images, specifically
the SIFT detector.  I then parse through all of the modified images and calculate
their keypoints as well.  Then I used a brute force matching algorithm to
match the most similar SIFT detection points between the two images.  I then
take the 25 best matches and sum their distances.  The original image with the
smallest keypoint difference from the modified image is selected as a match.

This algorithm is pretty poor, and does not work in any kind of speed.  A lot
of my issues were actually from instantiating the libraries.  The script does
not perform in a reasonable amount of time.  However, I think that it is a strong
initial start.  I have also tried to make a testing script called test.py, to
analyze some of the statistics from the original code, but again had a ton
of issues with fully utilizing the library.  The code I have is very buggy,
and will not execute fully anyway.

One of the major issues was the incredible amount of time a O(N^2) algorithm
would take.  Obviously this does not scale at all.  Major improvements could
be made to this algorithm.  First would be limiting the total number of 
SIFT detection points, perhaps to 25.  Another change would be to change the way
the search algorithm works, by using a kmeans approach, and not having
to search through every single image.  Also in the original dictionary, there
is no sense of images that could be out of the directory.  Instead, baseline
statistics would be needed, which I'd try to establish from the test.py in order
to known what kind of range of distance would be allowed

The libraries that I used were all from the python opencv, and based a lot on
an opencv tutorial website:

http://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_tutorials.html

I had used the SIFT algorithm before in MATLAB, not in python, so a lot
of the effort was put into the pushing the opencv library into an AWS 
machine.  This clearly was a lot of the struggles of my development today.
Specifically, I used the SIFT_detector and the brute force matcher for
my calculations.