[FFmpeg-devel] [RFC] H.264 error concealment and mpegvideo frame handling

[Jason Garrett-Glaser]

On Fri, Sep 24, 2010 at 10:56 AM, Michael Niedermayer <michaelni at gmx.at> wrote:
> On Wed, Sep 22, 2010 at 01:29:35AM -0700, Jason Garrett-Glaser wrote:
>> ffh264 does not handle the case of missing reference frames correctly.
>> ?Suppose we have 16 refs and we just decoded frame num 18. ?Here's our
>> reflist:
>>
>> 18 17 16 15 14 13 12 11 10 9 8 7 6 5 4
>>
>> Then, we decode frame num 20 (19 went missing):
>>
>> 4 18 17 16 15 14 13 12 11 10 9 8 7 6 5
>>
>> This is obviously buggy and wrong. ?So I do this to fix it:
>>
>> --- libavcodec/h264.c (revision 25148)
>> +++ libavcodec/h264.c (working copy)
>> @@ -1905,6 +1905,8 @@
>> ? ? ? ? ? ? ?s->current_picture_ptr->frame_num= h->prev_frame_num;
>> ? ? ? ? ? ? ?ff_generate_sliding_window_mmcos(h);
>> ? ? ? ? ? ? ?ff_h264_execute_ref_pic_marking(h, h->mmco, h->mmco_index);
>> + ? ? ? ? ? ?ff_copy_picture(h->short_ref[0], h->short_ref[1]);
>> + ? ? ? ? ? ?h->short_ref[0]->frame_num = h->prev_frame_num;
>> ? ? ? ? ?}
>>
>> ? ? ? ? ?/* See if we have a decoded first field looking for a pair... */
>>
>> This gives us:
>>
>> 18 18 17 16 15 14 13 12 11 10 9 8 7 6 5
>
> shouldnt that be 19 18 ... ?

19 doesn't exist.  We want to copy 18 into where 19 would have been,
and give it 19's frame number.

> either way ff_copy_picture() is wrong, we should for the entries that are
> left after gap handling allocate seperate pictures and fill them (this
> filling could then in the future be replaced by code that uses the b frame
> direct mode and motion vectors of the surrounding frames)

Attached is an extremely hacky and stupid patch that fixes the
problem.  Here's a sample that makes it blatantly obvious:
http://www.mediafire.com/?6huh99qemkrp11y .  This sample has a frame
dropped every 5 frames, then has the encoder fix it via reference
invalidation.

Dark Shikari
-------------- next part --------------
A non-text attachment was scrubbed...
Name: test2.diff
Type: application/octet-stream
Size: 1324 bytes
Desc: not available
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20100924/de33f7e5/attachment.obj>