[FFmpeg-devel] ffmpeg development priorities

[Neil Brown]

On Thursday August 16, aurel at gnuage.org wrote:
> 
> Indeed, if the above mentioned references are correct, ffmpeg should
> already be mostly ready to support AVCHD. My guess is that the mpeg-ts
> demuxer might need some small adjustments and that should be
> enough to support AVCHD.
> 

The AVCHD generated by my new Sony camera is 1080i and coded as
fields, not frames. i.e. a top field, then a bottom field, then a top
field ....

The h264 decoding in ffmpeg has some code to handle this, but it is
incomplete and buggy.

I have a patch (which you can find an early form of in the mailing
list archives) that fixes some bugs so that the first field (an
intra-coded TOP field) is decoded correctly and returned as the first
frame.  The second field (an inter-coded BOTTOM field coded in
reference to the TOP field) is almost correct, but not completely.
After that, decoding is a mess.

The issues that need to be addressed, as I understand them, are:

1/ Two consecutive fields need to be combined into a single frame
  before being returned to the caller.
2/ The storage of reference frames needs to understand that a frame
  can have two fields that are decoded separately but might be stored
  together.   A reference to a previous field might to the TOP field
  or the BOTTOM field of a previous frame.  It can even be to the TOP
  field of the current frame (when decoding a BOTTOM field).
3/ (I think) interpreting of motion vectors needs to deal with the
  possibility that the reference and current fields might of different
  parity (one a TOP field, one a BOTTOM field)

I think that once the correct approach to these issues are understood
and implemented, ffmpeg will do a much better job of AVCHD decoding.

NeilBrown