Constraining visual expectations using a grammar of scene events