home *** CD-ROM | disk | FTP | other *** search
- Tips on Encoding
- ----------------
-
- We have done some initial tests on encoding using different parameters.
- Here's a brief summary of what we've found, along with some suggestions.
-
- Subsample search gives almost identical compression to exhaustive search,
- and is somewhat faster (how much faster varies depending on other parameters).
- This seems to suggest that subsample search should almost always be used
- in lieu of exhaustive search. On the other hand, two-level search gives
- results similar to half-pixel exhaustive and subsample search, and is
- faster than both. So two-level search is a good alternative.
-
- All of those search routines give slightly better results than logarithmic
- search, but at a much higher speed cost. Which you use depends on how
- much time you're willing to trade off for compression.
-
- An important observation is that half-pixel search only increases the time
- for logarithmic search slightly, whereas it increases it dramatically for
- subsample/exhaustive search. Half pixel does give slight improvement in
- compression. Therefore, one rule of thumb is if you're using logarithmic
- search, it definitely makes sense to do a half-pixel search. On the other
- hand, if you're using subsample or exhaustive search, then it doesn't make
- sense unless you're willing to wait a long, long time -- a better alternative
- is to use two-level search (which does the exhaustive full-pixel search, then
- a local half-pixel search).
-
- Cross-2 B-frame search gives slightly better compression at the cost of
- roughly half the speed.
-
- The search technique with a good tradeoff of compression ratio/speed seems
- to be logarithmic search with cross-2 B-frame search, original frame as
- reference, half-pixel search, with a range of about 10 pixels. The range
- depends on the image, though, so you might want to experiment on a few
- representative frames before encoding the entire sequence. These settings
- are in the default parameters file.
-
- One important note about this discussion: All of the above is irrespective
- of quality. We are beginning to do experiments involving quality, and
- preliminary results seem to suggest that the most notable differences in
- quality occur between using original/decoded frames as reference frames, and
- between half/full pixel motion vectors. More detailed results will be
- forthcoming.
-