I'm not very impressed by the results because it shouldn't be too difficult to recognize the same image in slightly different positions. ARP is claimed to be robust to rotation, translation, and erosion yet they show in fig 4 the case where this doesn't hold. They don't show why q52 is a good query and what made q10 a poor search. I'm not very confident in the ablility of this method to actually do more complicated queries, aka a query with a different image than on from the database. Yet the retrieval results seem quite good. As for the speed constraint it does't seem to mean much for the video search criteria they reference. Based on other papers we've seen SIFT works quite well and this would probably be put in to some text retrieval framework so computation time isn't that important except for the 1 sketch query image. What I would be interested in seeing is a little more discussion about the binning/partioning with respect to orientation and maybe more importantly object size.