[FFmpeg-devel] [PATCH]: Fix decoding of mpegts streams with h264 video that does *NOT* have b frames

[Tony Strauss]

On Thu, Apr 21, 2011 at 3:34 PM, Mike Scheutzow
<mike.scheutzow at alcatel-lucent.com> wrote:
>
> Tony Strauss wrote:
>>>
>>> You can reproduce by running:
>>> ffprobe -loglevel debug -show_packets 720.no_b_frames.ts
>
> This .ts file does not look valid to me. I find no start codes in the video elementary stream. As far as I know, H264 should be using Annex-B format within a transport stream.

Hmm.  I don't have the command-line that I used to produce that file
anymore.  Perhaps I uploaded the wrong file.  Just to clear up any
doubts, I've uploaded some new .ts files, and I've provided the ffmpeg
command-lines that I used to create them from an .mp4 below:

The original mp4:
https://s3.amazonaws.com/tony-strauss-public/clip.mp4

A .ts file encoded with b-frames:
https://s3.amazonaws.com/tony-strauss-public/clip.bframes.ts
ffmpeg -y -t 5 -i clip.mp4 -an -vcodec libx264 -qcomp 0.6 -qmin 10
-qmax 51 -qdiff 4 -crf 18 -coder 1 -flags +loop -cmp chroma -g 48
-keyint_min 15 -sc_threshold 40 -rc_eq 'blurCplx^(1-qComp)' -i_qfactor
0.71 -partitions parti8x8+parti4x4+partp8x8+partb8x8 -trellis 1 -bf 1
-flags2 +wpred+mixed_refs-dct8x8-bpyramid -me_method umh -me_range 16
-subq 7 -refs 1 -threads 0 -f mpegts clip.bframes.ts

A .ts file encoded without b-frames:
https://s3.amazonaws.com/tony-strauss-public/clip.no_bframes.ts
ffmpeg -y -t 5 -i clip.mp4 -an -vcodec libx264 -qcomp 0.6 -qmin 10
-qmax 51 -qdiff 4 -crf 18 -coder 1 -flags +loop -cmp chroma -g 48
-keyint_min 15 -sc_threshold 40 -rc_eq 'blurCplx^(1-qComp)' -i_qfactor
0.71 -partitions parti8x8+parti4x4+partp8x8+partb8x8 -trellis 1 -bf 0
-flags2 +wpred+mixed_refs-dct8x8-bpyramid -me_method umh -me_range 16
-subq 7 -refs 1 -threads 0 -f mpegts clip.no_bframes.ts

The only difference between these two command-lines is the -bf (1 for
b frames, 0 for no b frames).  I believe that ffmpeg automatically
takes care of the annex-b repackaging during a re-encode.  Certainly,
when I explicitly added the filter on the command-line, the output
file checksum didn't change:

https://s3.amazonaws.com/tony-strauss-public/clip.no_bframes_and_filter.ts
ffmpeg -y -t 5 -i clip.mp4 -vbsf h264_mp4toannexb -an -vcodec libx264
-qcomp 0.6 -qmin 10 -qmax 51 -qdiff 4 -crf 18 -coder 1 -flags +loop
-cmp chroma -g 48 -keyint_min 15 -sc_threshold 40 -rc_eq
'blurCplx^(1-qComp)' -i_qfactor 0.71 -partitions
parti8x8+parti4x4+partp8x8+partb8x8 -trellis 1 -bf 0 -flags2
+wpred+mixed_refs-dct8x8-bpyramid -me_method umh -me_range 16 -subq 7
-refs 1 -threads 0 -f mpegts clip.no_bframes2.ts

blackberry:~/sandbox tony$ sum clip.no_bframes.ts
41807 608 clip.no_bframes.ts
blackberry:~/sandbox tony$ sum clip.no_bframes2.ts
41807 608 clip.no_bframes2.ts

>
>>> The
>>> key thing is that this file was encoded without b frames, so pts ==
>>> dts for every frame.
>
> It requires more than "no B frames" for this statement to be true.
>
> Does the video contain P-frames? If it contains P-frames, then pts != dts. Please post a link to a valid file so we can see.

Interesting.  I really don't know much about video encoding, but this
surprises me.  Certainly a patched ffprobe shows pts == dts for
clip.no_bframes.ts (and this does have p frames, since refs=1).  Why
would the presence of p-frames change pts == dts?  Since p-frames only
depend on an earlier frame, can't they be presented as soon as they're
decoded?

Tony