Intriguing... Who would have thought that a miniture image called an epitome would produce such results. I really like how the masking (or segmenting) works with a reconstructed image. The algorithm really doesn't have to "detect" anything just look for a particular patch. So, like in figure 3, if the patch of the ground is found in the epitome, it is easy to mask (or segment) this when the image is reconstructed.

Although the paper doesn't talk about tracking much, I can imagine this could easily be extended for tracking. Especially if an object in a video sequence is being tracked, a single epitome might be sufficient to track the object. Of course, if tracked objects are entering and exiting the scene, the epitome would need to be updated but probably not often (like every frame). This could have some great advantages.

Oh, and I downloaded the matlab code and tried it out. It doesn't match the results from the paper but it produces similar results. If you don't have time to run the code, I've put the results up here:

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

I also like the idea of using the epitomic to represent and reconstruct the image, it seems that it will reduce lots of redundant information, especially for the continuous video sequence. But when there are new objects appear in the image or when the existing objects move out of the scene, can we just update parts of the original epitomic or we need to train a new one to match the new frame? I don't know to update the models, maybe EM algorithm could also apply here.

Appearance epitomes are certainly impressive in its reconstruction abilities.

The shape epitome seemed like other layered approach at first glance, but never underestimate the power of inference!

The mask image is interesting in that it holds the opacity. I normally think of mask images as binary images... To me, using opacity to allow blending of layers is somewhat a novel idea.

A quick question: what information is needed for image reconstruction? Can this approach be used in image compression?

Love to see the explanation how the inference work in the presentation.

The idea of using this for compression sounds interesting. I was curious how the epitomes would fare compared to other methods of greatly reducing the size of an image in a study with humans where the participants were asked to recognize an image...