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.