[FFmpeg-devel] [PATCH 13/18] h264_sei: parse the picture timing SEIs correctly

[James Almer]

On 3/23/2020 1:17 PM, Anton Khirnov wrote:
> Quoting Kieran Kunhya (2020-03-22 05:24:35)
>> On Fri, 13 Mar 2020 at 10:32, Anton Khirnov <anton at khirnov.net> wrote:
>>
>>> Those SEIs refer to the currently active SPS. However, since the SEI
>>> NALUs precede the coded picture data in the bitstream, the active SPS is
>>> in general not known when we are decoding the SEI.
>>>
>>
>> The spec disagrees with this statement (7.4.1.2.1 Order of sequence and
>> picture parameter set RBSPs and their activation). There are reasonably
>> clear rules about SPS activation.
>> (Whether real world bitstreams comply is a different issue of course).
> 
> I don't think you read that section carefully enough. What it says (in
> my interpretation) is that an SPS can be activated only by:
> - buffering period SEI

Looking at our Buffering Period SEI parsing code, it doesn't look like
we're activating the SPS it references, for that matter.

> - a PPS (which is itself activated by picture data)
> Picture timing SEI is neither of those.
> 
> Also see D.2.2 Note 1, which suggests exactly what I'm doing with this
> patch.
>