It's hard to draw a line between broad object categories (e.g. "bike","wine bottle") and more specific object classes (e.g. "hartley & zisserman book cover"). This method seems to be targeting the former, while vocab tree methods like Nister et al. seem to be targeting the latter. Do you think this method would scale to a large set of more specific object classes like book or cd covers, or is it more tuned to borad object categories?

Can you explain more how individual object hypotheses are extracted from an input image. I love how they have the ability to detect multiple objects (even with partial occlusion), but how do they group certain features with certain object hypotheses?

(Comment-Only)
As for "Fast/Realtime Multiple Objects Recognition", it seems that some crucial building blocks are unavoidable(so far we have seen two papers):

(a) generative method (vs. discriminative method)
(b) vocabulary/codebook
(c) Tree model

It's a trade off between recognition accuracy, speed and scale.

(comment) I like how they combine all the features from different objects into the same tree. Interesting how they chose RPG wielders as one of their main object categories! It does seem like an easier class to recognize than some others.

What is agglomerative clustering method? .. well I guess nevermind because I googled it and found:

http://fconyx.ncifcrf.gov/~lukeb/agclust.html

I think it's interesting that this paper was funded by a project called CoSy. If you use it with your research, maybe you can call it CoSy MoZi! =)