Apply "-accurate_seek" alike default for `ffplay`

[Arnaud]

Summary of the bug:
How to reproduce:

% ffplay -ss 5
ffmpeg version: ffplay version N-78958-g8c24523
built from latest git, on Ubuntu 14.04.3, on 2016-03-07

File is seek_not_working.mp4

The file plays well but always from the beginning: the seek isn't working.

[Arnaud]

(will upload sample file)

[Arnaud]

Problematic file uploaded on the ftp server upload.ffmpeg.org
I would like to help finding what is the issue and maybe fix it if I can, but I'm not expert with ffmpeg source code, so if someone more experienced has an idea about what is wrong :)

[Carl Eugen Hoyos]

Is the issue not reproducible with ffmpeg (the application)?

[Arnaud]

I didn't know ffmpeg has a seek option.
I tried it the following way:
./ffmpeg -ss 3 -i seek_not_working.mp4 -frames:v 1 out1.jpg
and yes, it works.

[Carl Eugen Hoyos]

Did you find a sample with fourcc AVd1?

[Arnaud]

Sorry, I'm not sure I understand what you meant.
You're asking if I have the same problem with other mp4 files?

[Carl Eugen Hoyos]

No, you promised me in September to search for such a sample. Did you find one?

[Arnaud]

Ah, ok.
No sorry: I probably lost it when my hard drive crashed some time ago :(

[Arnaud]

Did someone has any clue about what might be wrong?

[Carl Eugen Hoyos]

I suspect the file has only one recovery point.

[Arnaud]

Thanks for your reply.

How can I check if the file has only on recovery point and if that can be the issue?

I tried to figure this out by adding a breakpoints in h264.c:decode_recovery_point but never get there, using my "buggy" file, but also some other files (e.g. [1] from [2]) for which seeking works.

[1]: ​http://www.sample-videos.com/video/mp4/720/big_buck_bunny_720p_50mb.mp4
[2]: ​http://www.sample-videos.com/

[Carl Eugen Hoyos]

Replying to Arnaud:

[1]: ​http://www.sample-videos.com/video/mp4/720/big_buck_bunny_720p_50mb.mp4

This file contains many IDR frames.
I don't know the exact semantic definition of recovery points but all IDR frames can be used for seeking.

[Arnaud]

Oh, OK.
Indeed that seems to be the problem in my initial file: only one IDR frame.
Then I guess this bug can be closed, because there isn't much to fix on ffmpeg library or on ffplay about it, don't you think?

If I'm right VLC uses ffmpeg, and I noticed it is able to seek this file, but it seems like it reads all the frames from the IDR frame to achieve this.
So I guess it's up to applications to implement such behavior on top of ffmpeg's seek if not enough IDR frames are present.

Thanks a lot for your help :)

[Carl Eugen Hoyos]

No IDR- (and no I-) frames are necessary to seek a valid H.264 stream, recovery points are sufficient. A stream analyzer is needed to find out if your stream does contain more than one recovery point.

[Balling]

It actually has 1 keyframe only, as second I frame is not a keyframe, it is not IDR (as nowadays avc includes I frames that are not keyframes, but not with ffprobe show_frames, you need -skip_frame nokey, see bug #8820).

​https://samples.ffmpeg.org/ffmpeg-bugs/trac/ticket5309/seek_not_working.mp4

Are we planning to implement decoding seek like said in #3083?

[MasterQuestionable]

͏    Codec container wrestling each other...
͏    ​https://www.google.com/search?hl=en&gl=ca&num=10&q=AVC|H.264|H264|x264+recovery-point
͏    ​https://stackoverflow.com/questions/75635925
͏    ​https://mpeg.chiariglione.org/tags/sei-messages
͏    .
͏    And the Open GOP (Group of Pictures) non-sense:
͏    That would be essentially identical as well-set ordinary (Closed) GOP.

͏    Ideally each GOP shall be an independent file:
͏    As in: ​https://github.com/MasterInQuestion/talk/discussions/3#discussioncomment-9376594

͏    And videos are not necessarily not meant to be viewed frame-by-frame. (thus any artifact unacceptable)

[Balling]

Nope.

B frame in one GOP is able to reference a frame in a different GOP. A GOP that contains any such outward-referencing frame is known as an "open GOP".

What is worse: B frames can be references. Even each B frame can be reference frame.

[Balling]

So the bug is only in ffplay. Actually.

[MasterQuestionable]

͏    Essentially such referencing could be still represented with Closed GOP...
͏    Note the final referencing is certain: not changing randomly.

͏    Probably the Open GOP design is to support certain really unusual fringe cases:
͏    Where 1 GOP would like to reference part (not all) of other GOP.
͏    And using the big GOP containing all is somehow unacceptable.

͏    However there doesn't seem to be much valid cases for such usage.
͏    Plus the apparent havoc of general interop.

͏    See also: ​https://trac.ffmpeg.org/ticket/11055#comment:16

͏    ----

͏    Up: ​https://forum.videohelp.com/threads/396068-Determining-GOP-structure-open-or-closed#post2574433

͏    No wonder why lossless H.265 may give output bigger than H.264...
͏    ​https://www.hevcbook.de/hevc-picture-types/
͏    .
͏    Who may explain such sophistry?

͏    Low Efficiency Video Coding: ​https://trac.ffmpeg.org/ticket/11155#comment:2

͏    ----

͏    Current OGOP designs are essentially no different than 2x (or 3x) size ordinary (Closed) GOP:
͏    Only pointlessly more complex and troublesome.

[MasterQuestionable]

͏    Per <Seeking> # Input seeking.