Could you give a quick crash course on Kalman filters?

Thanks!

Could you go over recombination? In Section 5 they state that "eventually, adjacent neighboring blocks can be recombined to form larger blocks if found to have similar distributions" How often does recombination take place? Figure 3 shows it after the 4th iteration, which makes me believe it isn't a part of each iteration, but I'm not sure if that's just an example. Also, how are the iterations defined?

What happens when there is more than one person in the scene. Will each individual be tracked separately? Thank you!

very cool that the system learns as the amount of data increases.

re Anton: The Results section describes that their iterations are defined by # of objects tracked. I think their iterations denote when they split the blocks, but they do the recombination and link pruning throughout...because each iteration takes hours.
I'm curious how they organized their data structures - there's a lot of complexity in their technique and it needs to be running real-time.

Can someone explain how they defined 'accuracy'? A jump from 55% to 81% is certainly large, but I'm not really sure what they're measuring or if 81% is even close to being useful.

It is interesting that this type of paper comes out from the UK. London probably has more surveillance cameras per capita than any other big city. They monitor the underground, streets, and tax cars automatically entering the city centre.

Big brother is watching...

It seems like they got a lot of mileage out of using a very simple appearance model for objects (color histograms). One of the reasons they use these histograms is that they are invariant to position. I wonder if a SIFT like feature might provide more discriminative power between objects and also some degree of pose independence. On the otherhand given the problem setting such a feature might be overkill.

Edited by author 10-26-2006 04:15 PM
to Paul: There exist CSIFT:A SIFT Descriptor with Color Invariant Characteristics - For some reason is not that popular..maybe because adding color made the computation a lot heavier and the robustness gained is not that much compared with the traditional SIFT.

About the paper: I wonder how this would change if the backgrounds the corridors have constant traffic of people and maybe a bit more cluttered background?. Color histograms and CCCM seems to look pretty well using the links on this uniform corridors.

I wonder if their system would work better if instead of using ordinary surveillance cameras, they used one of the new "dual sensor" cameras that combines visible imaging and thermal imaging into the same unit. Then they could not only track objects using color information, but would have a temperature signature too. I think this could improve their system's tracking ability for tracking objects that emit heat (like people). Plus it would work at night. I don't know how well thermal imaging works outdoors or in a brightly lit office environment though. Anybody know anything about this?

Edited by author 10-26-2006 04:26 PM
Yet another application that gets by with a crappy (color histogram) object recognition system that could probably benefit from having layers of classifiers with histograms on top. I'm pretty curious why they found that RGB was the nicest color space for their histograms. That surprises me somewhat, is there some nice intuition as to why that would be the case? (sensor sensitivity maybe?)

In response to the big brother: most survailence cams to the best of my knowledge are still black and white so we might be safe for another couple of years.

Somewhat related to my previous question: how does this method actually stack up with calibrated/supervised/overlapping-view methods? The amount of ad-hoc setup this method allows is certainly nice, but how much are we giving up?