Joelgranados
commented
12 years ago I'm not too sure about this one. I expect for a series of pictures to be located within a sub-region of the total amount of colors. That is, there is no reason to expect all the colors to show up in the analysis. So looking at the colors that are not included in the model might be overkill. Said in another way: Some colors will not be included in the model, not because the model needs to be retrained to include these colors; but because the series of pictures does not contain these colors. This reflection is even more relevant in color spaces like CIELUV and CIELAB where the distributions of colors is more compact.
When training the algorithm there is a probability of a bin not getting full by foreground nor background. This is not necessarily bad but the classification (training) of that bin as foreground or background can increase the accuracy of the model. The model calculation should specify this when calculating the model.
It could say something like: bins (10, 140 and 30) where not contained in foreground nor background. It should also give the ranges represented by each bin and should be capable of showing the user where these pixels are located so the user can decide to train them as foreground or background.