The paper does a good job of describing the features used for classification and how they are computed, but the choice of classifier seemed to be rather arbitrary. Some discussion on why this particular classifier was chosen and/or evaluation of different classifiers would probably have been a good idea. Also, the dataset used for training/testing was not described in any detail (and as far as I can tell it is not available on the web) so that it is hard to evaluate how good the classifier really is.

I found the idea presented in the paper useful and interesting, but i agree with the Markus in that that paper did not provide enough details. Furthermore i think that the two GUI tools that were developed were not so interesting. Although i have not installed them, i assume they are nice little toys, but i question their usability. The features used where well defined in the paper, but the way these were selected was neglected. I think it would rather be interesting to have a system that automatically selects some features; later, these selected features could be analyzed.

so you pick up a bunch of features and do something with them.
I wish not to be so harsh but that is what the paper is. There is little justification for why a particular feature was chosen over other choices (other choices are not even mentioned). The results look impressive, but then what is the baseline ? how hard is the problem ? Is mixture of gaussians the best you can do ? Probably not. I agree with Markus, the gui looks more like a toy rather than a serious content based retrival system.

I really like the topic that was presented in this paper. However I thought they should have talked more about the actual algorithms that were used for the genre categorization. Also, I thought that their explanation for why the best predicted genres were so - that they are less broad and their boundaries are less fuzzy, was tenuous at best.