SoC 2009/Weighted P-frame Prediction

From VideoLAN Wiki
Revision as of 06:23, 12 January 2010 by Underground78 (talk | contribs) (Vadalism Undo revision 15317 by Talk))
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search
This project is part of Google Summer of Code 2009.
Student: Dylan Yudaken
Mentor: Jason Garrett-Glaser


x264 is a highly popular h264 encoder. It currently does not implement the entire spec of h264. Currently weighted P-frames are not used to assist in prediction. If implemented this would give huge benefits in cases where the scene fades or where there are flashes. I am proposing to implement an implementation of this that is good enough and fast enough to warrant inclusion in most video encodings.


My patches (and codebase) are here for now, but they may soon move to videolan's server


No. Task Description Status Comment
1 Build weighting framework done .
2 Brute Force Optimal Weights done
3 Try use these new weights for some interesting things to improve quality done
4 Find Optimal Weights in a fast way (probably based on the findings in 3.) done
5 DSP Functions Done
6 Try duplicating reference frames Done
7 Other weird uses done Implemented a k-means cluster algorithm to find a number (k) of good weights and use them
8 Finish off small things Working Todo: Finish multipass, cleanup k-means a bit
9 Optimize algorithms and test like crazy Not started
10 Cleanup and code review Not started

Current Issues


This is an estimation.

March 20th or so - April 3 Qualifying task. Was a first attempt at goal 1.

April 20 Students announce

April 21 - May Goal 1 redone better and acceptably.

May 1 - May 20 Work on goal 2

May 23 Program Starts

May 26 - June 12 Exams

June 13 - July 13 Code

July 13 Mid Term deadline. Goals for here have not been set but personally I would like to have goal 2 & 4 finished with goal 3 having a decent chunk of work done. This weird out of order thing is because goal 3 can be worked on continuously as it is not really set, more the trying of ideas.

July 13 - August 17 Code

August 17 Program end