Bag of Words: Histograms
The next step was create a function that would take an image, densely sample SIFT features, determine which
word corresponded to each SIFT feature, and create a histogram of the resulting words.
I initially performed the word assignment using a dot-product comparison. Since the words were not normalized,
this put an implicit weight (for comparison purposes, not for binning purposes) on the higher norm words.
This was, unfortunately, quite slow, so when it came time to move on to better things, I went ahead and
installed the kd-tree modification.
As stated in the notes, the kd-tree gave an appreciable speedup. Its 'performance' (in terms of 'accuracy') was not obviously better
or worse than the dot-product assignment above, despite the fact that the kd-tree minimizes the l2 norm
rather than looking at directions.