re: Fei-Fei and Perona

This approach seems very principled and appears to get decent results, although 64% is not astounding in itself. Even though results are reported for categorization success if the top two choices are included, I would be more interested in knowing a different statistic: how many levels must you climb in the dendrogram in order to interpret a categorization as successful? Although I am not clear on the pseudo-Euclidean distance metric used, I think this would provide an interesting counterpart to the confusion matrix in showing that miscategorizations tend to mistake indoor scenes for other indoor categories, for example.

I'm intrigued by the prominent peak in the plot of figure 10(c), showing number of codewords vs. performance. Given their model, I don't understand why additional codewords can decrease performance. It's interesting to see how figure 11(a) shows a definite ceiling in the number of codewords which can obtain significance.

It would be interesting to see the performance of this algorithm using a set of codewords chosen a priori rather than learned, such as those detected by the human visual system, or a commonly used filter bank.