todo

来自「Motion JPEG编解码器源代码」· 代码 · 共 98 行

IN PROGRESS

	- Remove obsolete i,j in macroblocks in favour of Coord....
	- Replace appending subsampled luminance to luminance data with a 5-plane
	  picture data model (YUVHQ)
	- MotionEst should be called 'MacroBlockCodingMode'

	- Set satlim and scale type in quantizer during init

	- Move to a buffering output model where a frame can be aborted/restarted
	with different quantisation if it is coming out too large or too small.

	- Dynamic B/P selection and selection of I frames 
	without memory-expensive look-ahead.

	- Improve memory management.  Selecting Picture's from a single pool
	rather than seperate B / I|P pools will save memory buffers.  Also,
	really there is no need for DCT blocks for every picture as encoding
	is basically serial!!


Near Future improvements

0.  Replace current variance based motion estimation mode selection
	with DCT based selection.  
	Next Step: split predictors from Picture state and build a
	new QuantiseAndPutEncoding that predict's and transforms and
	quantisatises forall the motion estimates and selects the one that
	worked best for encoding.

	- Modify the on-the-fly rate controller so it can handle adjusting
	  quantisation at select Macroblock's and not at *every* macroblock
	  if no fancy per-block quantisation fiddles are active...

1. Add single coefficient dropping to DCT...

2.	Split into front-end and library (for gstreamer etc) 
	This will require removal of lots of nasty global scratch variables

3.	Add accurate (GOP-structure based) decoder buffer modelling for more
	aggresive bit-rate optimisation.

4.	Seperate the transformation phase (ME, pred, DCT) from the bit allocation
	and coding phases. 
	This will provide the basis for two-pass encoding and pipelined encoding.

5.	Replace the current frame-based parallelism with striping in the 
	ME/pred/DCT routine.

6.	The rate control code is now a mess of int64_t , double and int. Needs
	tidying... 

Minor Tidying

1.	A lot of parameter passing could be removed from motionest.cc routines
	by creating an 'MacroBlockMotionModes' object carrying parameters fixed
	for a block...

2.	The order of return from block_dist1_MMXE is silly.  Why not
	retune in the order it is held and adjust find_best_one_pel_MMXE to
	suit.  

3.	For Motion estimation in -I 1 we really need both field *and* frame
	sub-sampled images for motion estimation to be optimal.


5.	The code for scene-change gop sizing and ending sequences can in
	some circumstances mess up as the sizing takes no account of
	the fact the when a sequence split occurs the I-frame tempref is
	0 not M-1.


Sometime....



1.  	Bring field sequence MC/prediction to same level as progressive	
	sequence.

	
2.	Improve locality of reference by motion-estimating entire 4x4, 2x2, 1x1
phases in a single pass rather doing each for each Macroblock.

3.  Check out using dist2 for initial MC calculations as well as final
	selection?


5.	Add low-passing for high-activity regions when a specified quantisation
	level is exceeded (or block boundary edge energy exceeds some threshold?).

6.	Sort out that wretched vbv verification code.

	The mblockq_ sub-routines intended specification should be documented...