1 A Quick Description Of Rate Distortion Theory.

3 We want to encode a video, picture or piece of music optimally. What does

4 "optimally" really mean? It means that we want to get the best quality at a

5 given filesize OR we want to get the smallest filesize at a given quality

6 (in practice, these 2 goals are usually the same).

8 Solving this directly is not practical; trying all byte sequences 1

9 megabyte in length and selecting the "best looking" sequence will yield

10 256^1000000 cases to try.

12 But first, a word about quality, which is also called distortion.

13 Distortion can be quantified by almost any quality measurement one chooses.

14 Commonly, the sum of squared differences is used but more complex methods

15 that consider psychovisual effects can be used as well. It makes no

16 difference in this discussion.

19 First step: that rate distortion factor called lambda...

20 Let's consider the problem of minimizing:

22 distortion + lambda*rate

24 rate is the filesize

25 distortion is the quality

26 lambda is a fixed value choosen as a tradeoff between quality and filesize

27 Is this equivalent to finding the best quality for a given max

28 filesize? The answer is yes. For each filesize limit there is some lambda

29 factor for which minimizing above will get you the best quality (using your

30 chosen quality measurement) at the desired (or lower) filesize.

33 Second step: splitting the problem.

34 Directly splitting the problem of finding the best quality at a given

35 filesize is hard because we do not know how many bits from the total

36 filesize should be allocated to each of the subproblems. But the formula

37 from above:

39 distortion + lambda*rate

41 can be trivially split. Consider:

43 (distortion0 + distortion1) + lambda*(rate0 + rate1)

45 This creates a problem made of 2 independent subproblems. The subproblems

46 might be 2 16x16 macroblocks in a frame of 32x16 size. To minimize:

48 (distortion0 + distortion1) + lambda*(rate0 + rate1)

50 we just have to minimize:

52 distortion0 + lambda*rate0

54 and

56 distortion1 + lambda*rate1

58 I.e, the 2 problems can be solved independently.

60 Author: Michael Niedermayer

61 Copyright: LGPL