SoC 2009/Weighted P-frame Prediction
This project is part of Google Summer of Code 2009.
|
Contents
Abstract
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.
Code
My patches (and codebase) are here for now, but they may soon move to videolan's server
Goals
| No. | Task Description | Status | Comment |
|---|---|---|---|
| 1 | Build weighting framework | done | I have tested it to work. |
| 2 | Brute Force Optimal Weights | Mostly Working | Working well in hires, quite well in lowres. see below |
| 3 | Try use these new weights for some interesting things to improve quality | Not started | |
| 4 | Find Optimal Weights in a fast way (probably based on the findings in 3.) | Started | Lots of work to do here |
| 5 | Optimize | Currently working on the weight assembler code. | Currently mmx & sse2 functions work - but code needs to be done in a better way |
| 6 | See if explicit B-frame weighting can do good things | Not Started | |
| 7 | Other weird uses | Not Started |
Current Issues
- Currently the weighting doesnt work as well when looking at the lowres frame as the hires. This is the biggest issue currently and will be looked at after exams.
Timeline
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
