Isn't this a similar problem to the blurry-face vs. leaf-face issue discussed? If yes, then one would hand it off to computation to figure out which parts matter and need doctoring, except scaled up to video.