functional testing an image model

Frean and I—hacking for the fourth day in a row at the Victoria University Vic Books location—started the day by understanding, in real and fake data, how our image model likelihood behaves as a function of position and region size. There were some surprises there, all obvious in retrospect: One was that if there is any gradient or non-stationarity in the background pixels, the model wants to split the background into two (or more) huge boxes. Another was that the success of the model in distinguishing source from background is not strongly dependent on the prior we put on the background histogram model. That is, even if we give the same prior on histograms to the background and the source, the model still separates them.

By the end of the day we had a rudimentary optimizer and sampler for our model. Even with all our index trickery the model evaluations are still expensive. We hope to build a greedy optimizer version called "OExtractor" as a swap-in replacement for you-know-what!

No comments:

Post a Comment