HEVC Conformance Test Failed

Previous Topic Next Topic
 
classic Classic list List threaded Threaded
6 messages Options
Reply | Threaded
Open this post in threaded view
|

HEVC Conformance Test Failed

Zhenan Lin
Hi, experts & developers:

We use FFmpeg 3.4 to decode HEVC conformance bitstreams (
http://wftp3.itu.int/av-arch/jctvc-site/bitstream_exchange/draft_conformance/HEVC_v1/)
and only to find that 4 bitstreams got error ouput and 1 bitstream failed
to decoded.

The command as follows:
ffmpeg -threads 1 -cpuflags 0 -i <INPUT> -f md5 -

The output md5 is compared to reference software HM 16.9(
http://www.hevc.info/)'s and expected to be the same.

CONFWIN_A_Sony_1/CONFWIN_A_Sony_1/CONFWIN_A_Sony_1.bit
HM 16.9: a3ce3f936ff69ff1ec2621a622dd37ac
FFmpeg: c0a13e81b3a68c4263f240eb99a281b0

SAODBLK_A_MainConcept_4/SAODBLK_A_MainConcept_4/SAODBLK_A_MainConcept_4.bin
HM 16.9: b9c98937bcb6be445348fdc89c208f1e
FFmpeg: 23f03fffa670f7d770302c819b04f439

SAODBLK_B_MainConcept_4/SAODBLK_B_MainConcept_4/SAODBLK_B_MainConcept_4.bin
HM 16.9: 031043495e9b9f4ce120afa316aa5705
FFmpeg: de53c4f4adb77faca65c994c5fa7e7d0

VPSSPSPPS_A_MainConcept_1/VPSSPSPPS_A_MainConcept_1.bin
HM 16.9: 1ddf74263cb4953cfdfcf99c563d88ea
FFmpeg: b379cff14efe998a82c9dbc3c8399a67

TSUNEQBD_A_MAIN10_Technicolor_2/TSUNEQBD_A_MAIN10_Technicolor_2.bit
HM 16.9: 1bed23df7253170eeb8efc3fd1b740a3
FFmpeg: N/A
_______________________________________________
ffmpeg-user mailing list
[hidden email]
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

To unsubscribe, visit link above, or email
[hidden email] with subject "unsubscribe".
Reply | Threaded
Open this post in threaded view
|

Re: HEVC Conformance Test Failed

Moritz Barsnick
On Wed, Nov 29, 2017 at 18:19:03 +0800, Zhenan Lin wrote:
> ouput and 1 bitstream failed to decoded.
[...]
> TSUNEQBD_A_MAIN10_Technicolor_2/TSUNEQBD_A_MAIN10_Technicolor_2.bit
> HM 16.9: 1bed23df7253170eeb8efc3fd1b740a3
> FFmpeg: N/A

This is obviously not supported, as ffmpeg reports:
[hevc @ 0xa3919a0] Luma bit depth (10) is different from chroma bit depth (9), this is unsupported.

The file's description says that this is not a misinterpretation by
ffmpeg, but an intent of the file:

  purpose: Check Transform Skip intra and inter with unequal bitdepth (luma: 10-bit, chroma: 9-bit)

I'm certain that a patch for ffmpeg's hevc decoder is welcome. ;-)

I can't comment on the MD5 mismatches. I understand the samples are
quite "special", and apparently hit some shortcomings of the
decoder(?).

Do all the other samples decode correctly?

Moritz
_______________________________________________
ffmpeg-user mailing list
[hidden email]
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

To unsubscribe, visit link above, or email
[hidden email] with subject "unsubscribe".
Reply | Threaded
Open this post in threaded view
|

Re: HEVC Conformance Test Failed

Carl Eugen Hoyos-2
2017-11-29 12:15 GMT+01:00 Moritz Barsnick <[hidden email]>:

> I'm certain that a patch for ffmpeg's hevc decoder is welcome. ;-)

(I am not 100% sure of this: If the patch would introduce a speed
regression for non-insane samples, it will likely be rejected.)

Carl Eugen
_______________________________________________
ffmpeg-user mailing list
[hidden email]
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

To unsubscribe, visit link above, or email
[hidden email] with subject "unsubscribe".
Reply | Threaded
Open this post in threaded view
|

Re: HEVC Conformance Test Failed

Carl Eugen Hoyos-2
In reply to this post by Zhenan Lin
2017-11-29 11:19 GMT+01:00 Zhenan Lin <[hidden email]>:

> CONFWIN_A_Sony_1/CONFWIN_A_Sony_1/CONFWIN_A_Sony_1.bit
> HM 16.9: a3ce3f936ff69ff1ec2621a622dd37ac
> FFmpeg: c0a13e81b3a68c4263f240eb99a281b0

Needs "-vf crop=412:236:2:0" for bit-identical output.
Not sure if we want to support this, in any case, no
warning is shown, so there is a bug.

> SAODBLK_A_MainConcept_4/SAODBLK_A_MainConcept_4/SAODBLK_A_MainConcept_4.bin
> HM 16.9: b9c98937bcb6be445348fdc89c208f1e
> FFmpeg: 23f03fffa670f7d770302c819b04f439
>
> SAODBLK_B_MainConcept_4/SAODBLK_B_MainConcept_4/SAODBLK_B_MainConcept_4.bin
> HM 16.9: 031043495e9b9f4ce120afa316aa5705
> FFmpeg: de53c4f4adb77faca65c994c5fa7e7d0

This is a deblocking regression, I opened ticket #6882.

> VPSSPSPPS_A_MainConcept_1/VPSSPSPPS_A_MainConcept_1.bin
> HM 16.9: 1ddf74263cb4953cfdfcf99c563d88ea
> FFmpeg: b379cff14efe998a82c9dbc3c8399a67

Same as for the Technicolor sample: This looks insane and
is therefore unsupported until a real-world encoder
produces such files / patch welcome if no speed-regression
needed.

Carl Eugen
_______________________________________________
ffmpeg-user mailing list
[hidden email]
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

To unsubscribe, visit link above, or email
[hidden email] with subject "unsubscribe".
Reply | Threaded
Open this post in threaded view
|

Re: HEVC Conformance Test Failed

Mark Thompson
On 29/11/17 12:58, Carl Eugen Hoyos wrote:
> 2017-11-29 11:19 GMT+01:00 Zhenan Lin <[hidden email]>:
>
>> CONFWIN_A_Sony_1/CONFWIN_A_Sony_1/CONFWIN_A_Sony_1.bit
>> HM 16.9: a3ce3f936ff69ff1ec2621a622dd37ac
>> FFmpeg: c0a13e81b3a68c4263f240eb99a281b0
>
> Needs "-vf crop=412:236:2:0" for bit-identical output.
> Not sure if we want to support this, in any case, no
> warning is shown, so there is a bug.

"-flags unaligned" does this in the decoder (and is used in the FATE tests which include this file).

I think unaligned top/left cropping is also considered insane unless/until some real encoder actually uses it - this single conformance file is essentially the only instance of it seen in the wild.

- Mark
_______________________________________________
ffmpeg-user mailing list
[hidden email]
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

To unsubscribe, visit link above, or email
[hidden email] with subject "unsubscribe".
Reply | Threaded
Open this post in threaded view
|

Re: HEVC Conformance Test Failed

Carl Eugen Hoyos-2
2017-11-29 15:29 GMT+01:00 Mark Thompson <[hidden email]>:

> On 29/11/17 12:58, Carl Eugen Hoyos wrote:
>> 2017-11-29 11:19 GMT+01:00 Zhenan Lin <[hidden email]>:
>>
>>> CONFWIN_A_Sony_1/CONFWIN_A_Sony_1/CONFWIN_A_Sony_1.bit
>>> HM 16.9: a3ce3f936ff69ff1ec2621a622dd37ac
>>> FFmpeg: c0a13e81b3a68c4263f240eb99a281b0
>>
>> Needs "-vf crop=412:236:2:0" for bit-identical output.
>> Not sure if we want to support this, in any case, no
>> warning is shown, so there is a bug.
>
> "-flags unaligned" does this in the decoder (and is used in the
> FATE tests which include this file).

Then the only thing missing is a warning.

Thank you, Carl Eugen
_______________________________________________
ffmpeg-user mailing list
[hidden email]
http://ffmpeg.org/mailman/listinfo/ffmpeg-user

To unsubscribe, visit link above, or email
[hidden email] with subject "unsubscribe".