From the article:

Yu and her team had to analyze a torrent of fMRI data to identify from thousands of movie clips the 100 frames that most likely matched a given voxel activity pattern. They then “averaged” these shapes to yield the outline of what the subjects were seeing.