Everyone should read the 7th and 8th articles from the Fergus paper. They are about feature extraction. Jenny and Clifford might be coding up some of the algorithms presented in those papers.
After Jenny and Clifford left, we discussed how we need to deal with feature extraction for our project. Sal's ideas are to have things like a window finder and a window-patch finder for the buildings. He thinks that we need to put together finders such as those for the different types of objects that we would like to classify. He thinks that a parts analysis is more feasible than the sort of more semantic approach Linda wanted to take. I think I should meet with Linda and discuss it with her too. I don't think the semantic approach is totally unfeasible yet...we'll see.When using probabilities to detect features, you have to deal with a lot of dimensions generally. The Fergus paper deals only with classes of objects that have very distinct features. I think our goal is to be able to deal with classes of less distinctive objects.
Generally, though, the Fergus paper extracts many of the different dimensions, does a principle component analysis to take down the number of dimensions, and then using maximum likelihood estimation techniques to pick out salient features.