Created attachment 266808 [details] [review]
Patch to adjust the weighting of videoconvert's scoring mechanism to preserve colour information

Currently the scoring system used by videoconvert to decide the most appropriate format conversion has a very low value for the loss of colour (a weighting of just 2), whereas the weighting for loss of depth information is considerably higher (a weighting of 32). This can result in the videoconvert element discarding all colour information and using a grayscale format when this format provides a higher bit depth per channel.

For example, when converting I422_10LE into something that videoscale can deal with the current weightings result in GRAY16_BE being selected.

This can be seen in action by executing the following pipeline:

gst-launch-1.0 videotestsrc ! video/x-raw,format=I422_10LE ! videoconvert ! videoscale ! videoconvert ! autovideosink

Here a videotestsrc creates a colour image using the I422_10LE format, however this then gets converted to GRAY16_BE before being processed by the videoscale element, finally it gets converted into YUY2 for display (this may vary depending on chosen sink).

Attached is an example patch which increases the weighting of colour loss over depth loss and so preserves colour information under these circumstances. I'm not sure if a value quite so high as I've provided in the patch is strictly necessary, I'm not especially familiar with the various factors being balanced so it might be a good idea if someone with more experience of videoconvert's scoring system could have a look into the issue.

Perhaps the scoring system should be reworked to look at depth loss over all channels, rather than depth per channel. GRAY16 may have a higher depth per channel than I422_10LE (16 bit vs 10 bit), but it only has one channel, so is discarding far more information than I422_10LE holds overall (and discarding more information than would be kept in a conversion to a 3 channel 8 bit format).

Thanks,
 Mike Sheldon.