Archived from groups: alt.video.digital-tv (
More info?)
Doug McDonald wrote:
> Stephen Neal wrote:
>
>>
>> Where frame rate conversion is required (say from 60i/p to 50i/p)
>> standards convesion is improving now to the point where 50->60 and
>> 60->50 conversion is more transparent than MPEG2 encoding... (i.e.
>> more motion artefacts are introduced by the MPEG process than the
>> conversion) Using phase correlation techniques especially means it
>> is not a major issue.
>
> I don;t understand how this conversion can be done well. Consider
> the typical scenario when I watch a 50->60 Hz conversion: A F1 car
> is travelling down the track to 100 mph, with buildings and fences
> with narrow highly visible poles moving rapidly, as the camera
> pans to keep teh car centered. How do you fix up the 5:6 motion of
> those highly visible discrete poles? Aboyt 90% of 50Hz originaled TV
> is of this sort. It looks bad.
I'm not an expert in current standards conversion - but I did to a degree
project 10 years ago on motion detection to remove film weave from archive
sequences - very crude stuff for my degree but I read around a lot and have
retained an interest. (I spent a while as a broadcast TV R&D Engineer before
a radical career change took me to a different career in TV... I know say
"Run VT" a lot...)
Early 50/60 standards conversions didn't do any motion detection much at
all, they either ditched or repeated fields / frames, and scaled, causing
horrid juddery movement - this is often still used on consumer grade kit I
believe (such as digital conversion VHS VCRs etc.).
Then 4-field/4-line adaptive conversion - which was a mainstay for many
years (and is still widely used in cheaper, news-grade scenarios) was
introduced. This uses varying amounts of picture information from 4
successive fields, across 4 lines to generate an intermediate picture. The
weighting between the 4-fields contribution is altered depending upon the
amount of total movement within a frame detected. This crudely decides
whether to average a lot across 4 fields or not - depending upon the total
amount of motion between the fields. It is a kind of blanket motion
detection - not done on an area, by area basis.
More sophisticated block-based motion detection was then introduced (similar
to the motion detection used to remove temporal redundancy in MPEG2
encoding) - allowing small areas of the picture to be compared
field-by-field/frame-by-frame to see how much they have moved in between the
source fields. This movement detection then allows the block movement to be
interpolated to create wholly new frames for the standards converted
output - estimated where the blocks of picture would have been if they had
been captured in the output field/frame rate. This is commonly done using
block matching, where an 8x8 or 16x16 picture area is compared with a
similar area in frames/fields either side of the current field/frame - and
the nearest match position is decided to be where that area was previously /
will move to consequently. This works for a lot of movement - but only
allows one motion vector per block (or even per pixel) - which falls over if
you have semi-transparency (say a transparent rolling caption over a panning
shot - where there is two lots of movement in the same block.) It also
suffers with detecting rotational movement - say a spinning wheel.
Snell & Wilcox latest conversion (based on BBC R&D, and used for their
high-end converters - the Platinum and the fully loaded Alchemist - not all
Alchemists have it?) is Phase Correlation. This uses maths which I'm not
sure I have ever understood to provide a far better motion analysis, and
provide far better interpolation of intermediate frames/fields. It copes
better with multiple layers of motion, as well as improving on rotational
motion detection. It is the best conversion available I am advised.
From your example - a high end converter will isolate the poles and the car,
detect that they are moving at different amounts between frames, as well as
detecting the overall motion of the camera. It will then do its best to
create intermediate frames based on the source picture information and the
motion vectors - so will detect the motion of the poles, and create frames
with them in positions they never were in in the source material. This only
works if the poles are correctly tracked - and can go wrong on odd movement
of areas of uniform texture - though in these cases the motion tracking
algorithms usually give up gracefully and the older blending of fields takes
place (which is less of a problem on uniform stuff)
Fomula One material may not be a good choice - I don't know how much control
broadcasters have over the conversion of this stuff - as much of the
coverage is now provided by F1 themselves?
In comparison an older, or current cheaper, converter, characterised by
blurred/juddery output, would just mix varying amounts of the source fields
together to create the output frames, and you'd get multiple imaging etc.
The easiest way to tell which kind of converter is in use in 50Hz land is to
see how much clean field-rate motion is present. The older converters blend
and blur so much that most of the field-based motion is averaged out to a
kind of blurry mush, whereas the newer ones (especially PhC based stuff)
gives very clear and fluid field-based motion.
I am approaching this as a 50Hz viewer watching material converted from
60Hz - so we are throwing away temporal information rather than making it up
though - so we probably benefit in that way, though when 576/50 stuff is
converted to 480/60 you have the benefit of sharper source material?
Steve