Daily CVPR - Tuesday

Katie Bouman (MIT) works with Michael Johnson, Daniel Zoran, Vincent Fish, Sheperd Doeleman and William Freeman on Computational Imaging for VLBI Image Reconstruction. So instead of building an earth-sized telescope, what they do is putting telescopes around the world. By collecting data from them simultaneously, they can get some information about the structure of the black hole. Then they can use this information to try to reconstruct the underlying image for that black hole. So what happens is for every two telescopes in the telescope array, they get a single measurement of the underlying images’ 2D spatial frequency. What happens is that earth rotation entails different measurements of these spatial frequencies. Then the problem reduces down to: what image would cause those measurements in the frequency plane, but also looks like an image that we expect to see? This is an ill-posed problem, Katie says: we can abstract away all of the astrophysics and just look at this as purely a computational imaging problem: what image would cause those measurements, but also kind of looks like what we expect images to look like? They go through some rigorous modelling of the system and understanding the different noise that creeps into the problem. One of those being atmospheric noise that really scrambles our signal and makes it very difficult to reconstruct. They try to account for it by adding a trick called phase closure directly into the imaging. They also introduced this idea of using patch priors to try to test the invariance of the imaging assumptions in the final image which is reconstructed. So using a bunch of different sets of images, they extract little patches from them, and learn a model for what those patches look like. Then they can use that model to try to push image optimization in a direction that favors those kind of features. Katie would like to see that under the different patch models, they get a very similar image. This would also helps make them confident in the final reconstruction that they eventually put out for what a black hole looks like. The team wants to make sure that what they are injecting isn’t going to bias them too much in their final results. 6 CVPR Daily: Tuesday Highlights Katie works with a group called the Event Horizon Telescope that is trying to take the first image of a black hole. But since a black hole is so compact and so far away, we would actually need an earth-sized telescope to be able to resolve structure on the scale of a black hole's event horizon.