Based on the average analysis, the apparent winners are the OPP/INT/SP and OPP/INT. From the analysis it is clear that the behavior of an OPP histogram outperforms that of the RGB histogram. OPP/INT/SP is the clear winner for highest average precision at lower average recall (where its lowest recall is ~60% at ~57% precision). OPP/INT performs 2nd best overall, and seems to provide better recall at lower precisions than other configurations, yet converges to a lower recall as the precision increases. This may have to do with the small size of the image database we're using, or be an artifact of the type of results we are looking for (context/object + spatial distribution).

I completed the naive google image rebuffing using a python wrapper for the Retrieve executable. It scrapes google image search (given a query and a max number of images to find) and reorders them according to the group average using OPP/INT/SP by default. Instructions for use are in README.txt