gives a random order of frames (samples?)
(1) achieve better encoding
(2) achieve a particular false alarm rate.
One way to evaluate this is to pick up a 'leave-one-out' frame. Test pixels in the frame against the nearest codeword. If accepted, calculate a RMS distance over the pixels on that frame. Otherwise, count 'not-accepted' to get a false alarm rate. Here, I got a question. Does the RMS distance represent 'color distance'? What about 'brightness distance'?