[FFmpeg-devel] AC3 Format Detection unreliable.

[Justin Ruggles]

Michael Niedermayer wrote:
> Hi
> 
> On Sat, Aug 04, 2007 at 08:33:03PM -0400, Justin Ruggles wrote:
>> Justin Ruggles wrote:
>>> Dan Sully wrote:
>>>> * Dan Sully shaped the electrons to say...
>>>>
>>>>>> Yes, please make the stream(s) available.
>>>> Also, I am on an amd64 system.
>>> Doesn't look like that's an issue.  I'm getting the same thing on my 
>>> system.  The file is valid.  The issue seems to be the frame size. 
>>> 640kbps 48kHz is the largest AC3 frame size, and I think it's not 
>>> fitting enough frames into the probe buffer to give a positive 
>>> detection.  I'll look into it.
>> The problem with this sample is two-fold.  The first issue is that the 
>> probe buffer in this case is only 2048 bytes.  The frame size for this 
>> sample (and any others which are 640kbps & 48kHz) is 2560.  Therefore, 
>> only 1 frame is detected.  I changed the threshold for a score of 50 
>> down to 2 frames, but this doesn't help here.  After I changed to 1 
>> frame (at the start of the stream) it now gets a score of 50, but then a 
>> 2nd issue occurs.
> 
> so you havent fixed the issue but you sucessfully broke the whole probe
> system in lavf, great
> 
> the ff_ac3_parse_header function checks only something like 18bits
> a single match at the start IS NOT enough for a score of 50%
> mpeg-ps checks for somthing like 60bits and does misdetect it so
> how many misdetections do you think you added by decreasing the
> bits for ac3 to 18 at the start or 36 at a random pos?
> 
> also keep in mind that if no detection happens in 2048 bytes lavf will
> increase the buffer size until a reasonable detection is achived!
> 
> so to summarize do NOT "fix" bugs if you do NOT FULLY understand why
> it doesnt work and how the code works, this comment is not intended
> to you but rather to alot of people here
> when i think of pts/dts issues old ogg-vorbis, flv, swf, ... patches
> which where applied these all added fantastically broken, huge, stupid and
> unneeded code to workaround bugs in some other part of the code

You're right, I apologize.  I should have tested it more with other 
formats before relaxing the detection criteria.  I'll change it back and 
try to find another way.

2048 bytes is enough for the 1st 5/8 of any AC3 frame, so the 1st CRC 
code could be used to detect a single frame with pretty close to 100% 
certainty.  Then you have the issue of AC3 inside a container.  But 
would it be okay to give a 50% score if the frame is at the start of the 
probe buffer and is a CRC match?

-Justin