Ah, back to Augmented Reality stuff! For the first paper, I think that focusing on planes is actually a pretty reasonable assumption. The vast majority of the time there will be at least two planes in which a homography can be determined. Anyway, the results are good but it bothers me that they must manually indicate a planar region in the first image of a video sequence. I'm sure that there is a way to automate this so that the planar region can be determined on the fly. By the way, I want to pick people's brains a bit. Can you think of a real life scenario (application) where this technique would break (other than the obvious of a camera facing a flat wall with no texture)?

The second paper is another important paper in the AR world (and I'm glad it is included here). The main thing is that they are able to determine the plane on the fly, but it requires "markers" with the two black polygonal regions that they use (see figure 5 and 14). This is very similar to what the ARToolkit does. So this is contrary to the first paper which was trying to remove markers altogether. Personally, I believe that there is a middle ground where you can determine a planar region automatically without the need for markers. This is currently where research in the AR world is headed.

One more quick thing... For the last paper regarding the epitomes, Shinko and I were wondering what an epitome of a color gradient would look like. Well I ran one through and the results are interesting. I think the result is a good example of how the patches are put together. Anyway, I've put the results up here:

http://www-cse.ucsd.edu/~mclothie/gradient_epitome.jpg

Just a brief introduction: the first paper presents a tracking algorithm based on planar structure. The second relax the tracking problem and concentrates on a novel way to represent object. As I had tried to implement both of these approaches, in my presentation, I will discuss some technical problems I have in the experiment.

Matt's question is an interesting one. One of the situations is when you have a lot of objects (people) moving around, because it will mess up the corner detection and corrupt RANSAC. Love to hear other ideas.

I've got a "fun" scenario where the plane technique will fail. For you Star Trek fans out there, imagine trying to do AR in outer space. If you have an image of just stars, there really is no way to determine a plane (especially since each star could exist in its own plane). Of course, I don't think this is something that we will have to worry about for a while. ;)

In the first paper, I'm wondering how well the "automatic plane detection" works. They argue that the best homography for the whole image will correspond to the homography for the largest plane in the image, but I'm not sure how robust this is. I'm also a little confused about how much hand initialization is going on. They say they specify the plane for the initial homography, but then they say something about setting "the boundary to the convex hull of the 2D points" for initialization...

And I guess my answer to Matt's question about where this will fail: scenes without planar surfaces, scenes where the planar surface has no texture (no corners will be detected), sequences where the initial plane moves out of view before a second plane becomes visible...

The outer space scenario is somewhat interesting. Since any translational motion will be miniscule compared to the distances between stars, all motion will be approximately rotational, meaning a homography will exist. The problem of course is that this doesn't really provide a frame of reference for synthesized objects since translation can't be detected.

I'm a little concerned about the initialization as well - it seems like you'd need to re-initialize every time the registered planar object moved out of the frame. Would I need to manually pick out something trackable to this algorithm every time I changed what I was looking at?

I have been pondering how is it that those folks at NFL made such a stable Yellow line for the first-down line. I guess this homoraphy tracking system is one way about it. Interested readers may like to know that htere's an article in teh Spectrum from IEEE last month regarding the awesome looking yellow line on the football field and teh 2 years and 2 million dollars they spent to create such a system.

Potentially, initialization can be done by having the camera face a perfectly planar surface. Using the same point correspondence algorithm over time, it can be determined automaitcally, via neighborhood correlation matching or some such measure, that the observed is planar, and the poitns can be taken as is to compute an initial H.

Need new Rip DVD to AVI ?
Rip DVD to AVI
Have a nice surfing！