[x264-devel] Re: Understanding Intra mode prediction

[Loren Merritt]

On Wed, 18 Apr 2007, Son Minh Tran wrote:

> I have some questions concerning how x264 selects the prediction mode for 
> intra coding
> -How did you derive the coefficient i_lambda=pow(2,QP/6-2)?

The reason for pow(2,qp/6): That's the function used to convert a qp to 
the actual number that dct coefficients are divided by.
The reason for the -2 (i.e. a constant factor of 1/4): I tried various 
constants, and that was the best one.

The process of determining the optimal lambda for a given qp is:
   foreach qp in [1..51]
     foreach lambda in [some list of candidates]
       encode the test video with this qp and lambda
   plot the bitrate and psnr of all encodes
   keep the lambdas that end up on the convex hull of the plot

The result is of course somewhat dependent on the content of the test 
videos and on the other options selected, so I don't claim that the 
lambdas I chose are always optimal.

> -For a flag b_fast_intra, I saw that  you  switch it  on  when   the block is 
> less likely  in intra mode. And you did nothing when the block is likely in 
> intra mode. Here is the extraction from the code:

In each macroblock: a->b_fast_intra is initialized to 0, may or may not be 
changed to 1, and either way the value is used later.

--Loren Merritt

-- 
This is the x264-devel mailing-list
To unsubscribe, go to: http://developers.videolan.org/lists.html