[3dem] MotionCor2 1.4.0 released

[Shawn Zheng]

MotionCor2 1.4.0 has made some important changes and implemented several
new functions, including support for EER (ThermoFisher) and optimization
for tilted data collection.  To support the increasing need for movies
collected at high tilt angles, MotionCor2 1.4.0 implemented a
distance-based scheme to interpolate and correct local motion at pixel
level. This strategy has been found more accurate than the polynomial
fitting used in earlier versions for tilted samples that are known for more
intensive local motion across the field of view.  An improved scheme has
been implemented that leads to more robust detection of failed local-motion
measurements on challenging patches. The failed measurements are then
replaced with values estimated from good ones in the neighborhood. This
improvement is beneficial not only for more accurate measurements but also
for finer measurements across the field of view.  MotionCor2 1.4.0 added a
new function that supports movies collected with variable frame exposure
that aims to freeze stronger early motions with shorter exposures and in
the meantime control the movie size with longer ones for the later frames.
Non-uniform grouping has therefore been implemented for this kind of movies
in which shorter-exposed frames are grouped more than those of
longer-exposed.  As a result, frames formed by summing each group on which
global and local motions are measured have an even distribution of
dose(signal). The motions of the original frames are then quadratically
interpolated.

The performance of MotionCor2 1.4.0 was kindly assessed by nVidia on their
platforms equipped with various advanced graphics cards. The first plot
shows the wall-clock time of processing a single TIFF movie containing 120
K3 super-res frames on a single GPU including time to load the TIFF movie
into CPU RAM. This plot is for users who are interested in on-the-fly
motion correction. We routinely run 8 jobs concurrently during the data
collection, one GPU per job on a workstation equipped with 8 GPUs and are
able to catch up the data collection speed.

The second plot is for offline processing. It shows the wall-clock time of
a single job using 1 and 2 GPUs, respectively to process 10 K3 movies
containing 120 frames in batch mode. In this case, the 10-second movie
loading time is shadowed by the computation of the movie in RAM.

If you have more questions regarding the performance assessment, please
feel free to let me know. Again, we really appreciate nVidia's various
kinds of help!

Best
Shawn
[image: SingleProcessing1GPUa.png]
[image: BatchProcessing2GPU.png]
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.ncmir.ucsd.edu/pipermail/3dem/attachments/20201101/f4c0047b/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: SingleProcessing1GPUa.png
Type: image/png
Size: 23964 bytes
Desc: not available
URL: <http://mail.ncmir.ucsd.edu/pipermail/3dem/attachments/20201101/f4c0047b/attachment-0002.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: BatchProcessing2GPU.png
Type: image/png
Size: 37443 bytes
Desc: not available
URL: <http://mail.ncmir.ucsd.edu/pipermail/3dem/attachments/20201101/f4c0047b/attachment-0003.png>