Bad pixel format negociation with scale and gray

[Clément Bœsch]

The gray pixel format seems to be preferred over colored output sometimes.

Working fine:

% ./ffplay -nostats -v verbose -f lavfi testsrc -vf 'split[a][b]; [a]pad=iw*2[x]; [b]edgedetect[y]; [x][y]overlay=w'       
ffplay version N-52968-g90d35e5 Copyright (c) 2003-2013 the FFmpeg developers
  built on May 11 2013 12:34:21 with gcc 4.8.0 (GCC) 20130502 (prerelease)
  configuration: --enable-gpl --enable-version3 --enable-fontconfig --enable-libfreetype --enable-libmp3lame --enable-libvorbis --enable-libxvid --enable-libx264 --enable-libvpx --enable-libtheora --enable-x11grab --enable-libopenjpeg --enable-libass --enable-libmodplug --enable-libv4l2 --cc=colorgcc --samples=/home/ubitux/fate-samples --prefix=/tmp/ffinstall --disable-runtime-cpudetect --enable-libcelt --enable-libopencv --enable-frei0r --enable-libcaca --enable-libiec61883 --enable-libopencore-amrwb --enable-libopencore-amrnb --enable-libopus --enable-libpulse --enable-libspeex --enable-libquvi --enable-memory-poisoning
  libavutil      52. 30.100 / 52. 30.100
  libavcodec     55.  7.100 / 55.  7.100
  libavformat    55.  4.101 / 55.  4.101
  libavdevice    55.  0.100 / 55.  0.100
  libavfilter     3. 64.100 /  3. 64.100
  libswscale      2.  3.100 /  2.  3.100
  libswresample   0. 17.102 /  0. 17.102
  libpostproc    52.  3.100 / 52.  3.100
[Parsed_testsrc_0 @ 0x7f84fc001600] size:320x240 rate:25/1 duration:-1.000000 sar:1/1
[ffplay_buffer @ 0x7f84f0000dc0] w:320 h:240 pixfmt:rgb24 tb:1/25 fr:25/1 sar:1/1 sws_param:
[auto-inserted scaler 0 @ 0x7f84f0006e80] w:0 h:0 flags:'4' interl:0
[Parsed_edgedetect_2 @ 0x7f84f0005880] auto-inserting filter 'auto-inserted scaler 0' between the filter 'Parsed_split_0' and the filter 'Parsed_edgedetect_2'
[auto-inserted scaler 1 @ 0x7f84f0002620] w:0 h:0 flags:'4' interl:0
[Parsed_overlay_3 @ 0x7f84f00060a0] auto-inserting filter 'auto-inserted scaler 1' between the filter 'Parsed_pad_1' and the filter 'Parsed_overlay_3'
[auto-inserted scaler 2 @ 0x7f84f00011c0] w:0 h:0 flags:'4' interl:0
[Parsed_overlay_3 @ 0x7f84f00060a0] auto-inserting filter 'auto-inserted scaler 2' between the filter 'Parsed_edgedetect_2' and the filter 'Parsed_overlay_3'
[Parsed_pad_1 @ 0x7f84f0004aa0] w:320 h:240 -> w:640 h:240 x:0 y:0 color:0x000000FF
[auto-inserted scaler 1 @ 0x7f84f0002620] w:640 h:240 fmt:rgb24 sar:1/1 -> w:640 h:240 fmt:yuv420p sar:1/1 flags:0x4
[auto-inserted scaler 0 @ 0x7f84f0006e80] w:320 h:240 fmt:rgb24 sar:1/1 -> w:320 h:240 fmt:gray sar:1/1 flags:0x4
[auto-inserted scaler 2 @ 0x7f84f00011c0] w:320 h:240 fmt:gray sar:1/1 -> w:320 h:240 fmt:yuva420p sar:1/1 flags:0x4
[Parsed_overlay_3 @ 0x7f84f00060a0] main w:640 h:240 fmt:yuv420p overlay w:320 h:240 fmt:yuva420p
[ffplay_crop @ 0x7f84f00030c0] w:640 h:240 sar:1/1 -> w:640 h:240 sar:1/1

Broken with a scale filter (the left is part is being "contaminated" with the gray-only:

% ./ffplay -nostats -v verbose -f lavfi testsrc -vf 'scale=iw:ih, split[a][b]; [a]pad=iw*2[x]; [b]edgedetect[y]; [x][y]overlay=w'                                                  
ffplay version N-52968-g90d35e5 Copyright (c) 2003-2013 the FFmpeg developers
  built on May 11 2013 12:34:21 with gcc 4.8.0 (GCC) 20130502 (prerelease)
  configuration: --enable-gpl --enable-version3 --enable-fontconfig --enable-libfreetype --enable-libmp3lame --enable-libvorbis --enable-libxvid --enable-libx264 --enable-libvpx --enable-libtheora --enable-x11grab --enable-libopenjpeg --enable-libass --enable-libmodplug --enable-libv4l2 --cc=colorgcc --samples=/home/ubitux/fate-samples --prefix=/tmp/ffinstall --disable-runtime-cpudetect --enable-libcelt --enable-libopencv --enable-frei0r --enable-libcaca --enable-libiec61883 --enable-libopencore-amrwb --enable-libopencore-amrnb --enable-libopus --enable-libpulse --enable-libspeex --enable-libquvi --enable-memory-poisoning
  libavutil      52. 30.100 / 52. 30.100
  libavcodec     55.  7.100 / 55.  7.100
  libavformat    55.  4.101 / 55.  4.101
  libavdevice    55.  0.100 / 55.  0.100
  libavfilter     3. 64.100 /  3. 64.100
  libswscale      2.  3.100 /  2.  3.100
  libswresample   0. 17.102 /  0. 17.102
  libpostproc    52.  3.100 / 52.  3.100
[Parsed_testsrc_0 @ 0x7fd048001600] size:320x240 rate:25/1 duration:-1.000000 sar:1/1
[ffplay_buffer @ 0x7fd04c000dc0] w:320 h:240 pixfmt:rgb24 tb:1/25 fr:25/1 sar:1/1 sws_param:
[Parsed_scale_0 @ 0x7fd04c004160] w:iw h:ih flags:'4' interl:0
[auto-inserted scaler 0 @ 0x7fd04c000b80] w:0 h:0 flags:'4' interl:0
[Parsed_overlay_4 @ 0x7fd04c007100] auto-inserting filter 'auto-inserted scaler 0' between the filter 'Parsed_pad_2' and the filter 'Parsed_overlay_4'
[auto-inserted scaler 1 @ 0x7fd04c006e00] w:0 h:0 flags:'4' interl:0
[Parsed_overlay_4 @ 0x7fd04c007100] auto-inserting filter 'auto-inserted scaler 1' between the filter 'Parsed_edgedetect_3' and the filter 'Parsed_overlay_4'
[Parsed_scale_0 @ 0x7fd04c004160] w:320 h:240 fmt:rgb24 sar:1/1 -> w:320 h:240 fmt:gray sar:1/1 flags:0x4
[Parsed_pad_2 @ 0x7fd04c005b40] w:320 h:240 -> w:640 h:240 x:0 y:0 color:0x000000FF
[auto-inserted scaler 0 @ 0x7fd04c000b80] w:640 h:240 fmt:gray sar:1/1 -> w:640 h:240 fmt:yuv420p sar:1/1 flags:0x4
[auto-inserted scaler 1 @ 0x7fd04c006e00] w:320 h:240 fmt:gray sar:1/1 -> w:320 h:240 fmt:yuva420p sar:1/1 flags:0x4
[Parsed_overlay_4 @ 0x7fd04c007100] main w:640 h:240 fmt:yuv420p overlay w:320 h:240 fmt:yuva420p
[ffplay_crop @ 0x7fd04c0030c0] w:640 h:240 sar:1/1 -> w:640 h:240 sar:1/1

[Clément Bœsch]

Note: this was originally reported in #2465.

[Carl Eugen Hoyos]

I don't think this is a valid ticket: FFmpeg tries to only insert the scale filter once and succeeds.

[Clément Bœsch]

Replying to cehoyos:

I don't think this is a valid ticket: FFmpeg tries to only insert the scale filter once and succeeds.

I see 2 auto-inserted filters, in addition to the explicit scale filter. And that behavior is not expected anyway; if a user wants to reduce the amount of auto-inserted filter, he can insert a format filter to pre-convert to gray.

[Cigaes]

cehoyos is right, the behaviour is normal.

The auto-inserted filters are on the pad-overlay and edgedetect-overlay links, they are completely unrelated to the problem at hand, which happens on the input side of edgedetect.

Note that you can get the same result without inserting the scale filter: just move the edgedetect definition before the split filter.

I do not think that "solving" this kind of problem is a practical solution. At some point, users need to realize that complex filter graph sometimes need explicit format specification.

[Clément Bœsch]

Replying to Cigaes:

cehoyos is right, the behaviour is normal.

The auto-inserted filters are on the pad-overlay and edgedetect-overlay links, they are completely unrelated to the problem at hand, which happens on the input side of edgedetect.

Note that you can get the same result without inserting the scale filter: just move the edgedetect definition before the split filter.

I do not think that "solving" this kind of problem is a practical solution. At some point, users need to realize that complex filter graph sometimes need explicit format specification.

While this behaviour might be "normal" with the current design, I don't understand how that can be the expected one.

How can a user know if he needs explicit format specification or no (except by observing random craziness in some cases)?

...and I still believe the "lossless" path should be prefered over the "optimized" one. Inserting a gray auto-scaler in this place "sometimes" is IMO a non-sense from a user experience PoV.

So while this might not be considered a bug in itself, it can be stated as a design problem or current limitation nevertheless. Feel free to reword and re-qualify the ticket.

[Michael Niedermayer]

IMHO one of multiple outputs of a split filter should not cause the other outputs to loose chroma or alpha but thats what happens. A similar issue could arise with "merge" filters