Log

Meeting Wed/ 11 Nov. 2008

Motivation: Want a very CPU-cheap initial ranking of categories, in order to send only a few categories to the classifier. This way we could scale visual recognition to 1M categories.
Need different way of looking at the results, find a tau for which, say, 90% are correctly classified.
- Goal: show that tau scales sub-linearly with the number of categories.
Think about the reviewers: need whole pipeline to convince the reviewers that it is not only fast, but that it also works very well in practice.