AI code related to the Gallery,
for example for predicting an image's popularity
(eg by vote or download count).
Image Popularity Prediction
There will be a number of levels of sophistication to popularity/"goodness" prediction;
lower levels will do simple parametric statistical analyses of 2D images
(possibly segmented into different types such as small-palette-based, true-colour and greyscale)
for example for the interquartile (or wider) distribution of properties such as
luminance (Y), hue (H), saturation (S) values or their frequency components
from sampling of (or sampling of thumbnails of) the images.
Parameters may be varied by genetic algorithm (GA) or other goal-seeking approaches.
Typically a number of sample points (16--256) will be chosen in a specific pattern
from the source image scaled to a 256x256 grid (retaining aspect ratio,
so for non-square images some points outside the scaled-down area will be "undefined").
Attempts will be made to handle multi-modal distributions of "good" values,
eg where if any (quorate count) of a number of analyses votes "for" an image then it may
be considered "good" (or bad if the votes are in the other direction).
Higher levels of sophistication may use more powerful techniques such as
generalised virtual machine code to do analysis.
Data Sparseness for Calibration
Given that calibration data is sparse,
ie few exhibits have explicit votes or frequent downloads or other external measures of "goodness",
calibration is likely to be confined to those exhibits with such data,
though secondary calibration against system-side synthetic "goodness" measures may also be used.