[theora] Why Our Civilization's Video Art and Culture is Threatened by the MPEG-LA

[Andrey Filippov]

In video recording there are additional requirements, different from
broadcasting/distribution -the goal is to preserve virtually all the
information  that the camera critical element (CCD or CMOS sensor) can
provide  and do some post-processing later. That "virtually all the
information" does not necessarily mean "raw" (uncompressed) image data -
reasonable compression that matches the noise performance of the imager
would not degrade the data significantly.

In our cameras we currently use MJPEG derivative (with standard JPEG
replaced with modified version
-http://community.elphel.com/jp4/jp4demo.php) with the goal of
postponing the demosaic processing of the color images to
the post-processing stage. Most color sensors provide color mosaic images
and modern algorithms are good at minimizing color artifacts, but it is a
separate function from video recording and when you do not need real time
video but rather want the best possible quality of the result (edited) video
it is better to record pixel values and only interpolate colors later,
combined with other editing.

We did have subset of the Ogg Theora encoder in the previous model (Elphel
333) and I agree - when ran in  keyframe-only mode it performed better
compression than JPEG, but the Ogg encapsulation introduced rather heavy
load on the small camera CPU. Theora encoding was done in FPGA, but CPU had
to calculate the frame checksums after combining hardware-compressed data
with the headers by the software. And as FPGA processing is about 100x
faster, that last step (Ogg encapsulation) was too expensive for most our
applications and in the current model 353 camera we unfortunately do not
have Theora, only MJPEG-JP4.

Our next camera under development has more processing power but the sensors
also grew in size since our Theora implementation in 2005. After the porting
of the current code we'll re-consider Theora and Dirac but we'll look from
our side - how they perform on the acquisition side (not the display), how
they address:
 1 - preserving the sensor data, keeping compression artifacts under the
noise floor;
 2 - convenience of the hardware (FPGA) implementation
 3 - compression ratio -only the last of the three requirements.


Andrey
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.xiph.org/pipermail/theora/attachments/20100505/721e220b/attachment-0001.htm