[FFmpeg-devel] Decoding artifacts with some AVCHD files

[Pascal Massimino]

  Gerald,

On Thu, Mar 20, 2008 at 2:31 AM, Gerald Dalley <gerald.dalley at gmail.com>
wrote:

> I have a Canon HG10 camcorder and I'm attempting to decode their AVCHD
> content.  Unfortunately, some frames don't seem to be decoding
> properly.  As an example, at several points (e.g. starting at about
> frame 52) of
>
>
> http://people.csail.mit.edu/dalleyg/collaboration/avchd/hg10-pf24-clip2.mts
>

 they are using field-based interlacing (aka PAFF) , with 2 bframes.
 Their top field for b-slice uses temporal bdirect mode, whereas
 their bottom field in b-slice uses spatial prediction bdirect mode.
 Support for this prediction is incomplete in ffmpeg, hence the artifacts
 you're seing.


> the bottom field image patches don't seem to be updated quite properly
> during a time when the camera moved quickly.  Using the InterVideo
> codec, the decoding seems to work fine (unless its always-on
> deinterlacer is erasing the mistakes) so I don't think it's a problem
> with the HG10's encoder.  Here's a direct link to a cutout frame
> demonstrating the problem:
>
>
> http://people.csail.mit.edu/dalleyg/collaboration/avchd/decoded/cropped0053bottom.png
>
> I see the problem using the latest clsid build of ffdshow on Windows
> from 17 March 2008 and using the latest SVN-built compilation
> of and ffmpeg on x64 Linux on 18 March 2008.
>
> I've written up an article that mostly discusses pulldown issues with
> the camera (which are not pertinent to this issue), but I also include
> this ffmpeg decoding issue.  Search for "Frame Corruption" in
>
>  http://people.csail.mit.edu/dalleyg/collaboration/avchd/readme.html


  very interesting article. Hell will freeze before
  we're rid of last millenium's interlacing.

skal