folks know where I get posses some Propecia ?

people have where I can posses some Proscar?

The cheapest provigil I have been able to find online is at www.24x7pharmacy.com which as 100 tabs of 200mg for $155. Incidentally they also have the lowest price on Soma - 100 tabs, 350mg for 26.99

it is good topic

Buying Provigil without a prescription is dangerous. It is always advisable to visit a physician before taking any prescription medication. To buy Provigil safely and have it delivered directly to your door, visit www.Discount-Canadian-Meds.com

My guess is that this approach is limited to sets of only a few object categories. As the number increases, it is increasingly difficult to visually categorize using a naive bag-of-words approach. I assume if the system was run on the COIL100 database with its many similar objects, or the ETH80 database with lots of intra-category variations, the results would be less impressive. Still, it would be interesting to know how far one could exploit this framework.

The topic discovery and classification applications demonstrate the utility of sift features and the probabilistic models, but fall short of being 'unsupervised'. I think that the most interesting bit from a practical point of view is the segmentation. Background images are easy to collect and could be used to train one of these to segment foreground from background. Then the foreground images could be fed into some more specific recognition framework. I guess this just adds up to a good tool for segmenting images - but it seems to be useful anyway.

I was thinking the exact same thing as Robin. As I was reading the paper everytime they mentioned how this was an unsupervised learning algorithm I had to think I missed something. Since they knew the total number of object classes, though they didn't individually label the objects, they were esentially separating those classes. I'm not sure exaclty what they used for the K-means vectors because it did significantly worse. But seeing K < 4 would be quite interesting. Even more to the point, in experiment 2 they actually classifiy on a separate set of 400 background images. So here they actually have 400 + 1 labels. In anycase this again speaks of the power of SIFT like features for recognition and benifits of adding some sort of proximity/geometry for image comparision.

The authors of course knew the underlying number of classes was 4. Since k=4 finds these classes, it's not surprising that larger k breaks out subclasses. What they don't tell us about is what happens if you set k=3. I imagine it's not pretty.