Project

General

Profile

RCT problems

Added by Anonymous over 14 years ago

Hi there!

I am experiencing problems with RCT and libCV (still Leginon 1.6):

When I try with a test grid, everything works quite well - the exposures at 0 and 30 degrees are shifted a little bit but there is an ~70% overlap. Correlation is done at HM 4700x (hl), because the en/sq alignment gets bad very fast. I do not mind that I miss a part of the square.

But with a real world negative-stain sample, the problems start:

The correlation does not work any more, although the hl images contain some clearly distinguishable features (dirt etc) but not as much as the test grid mentioned above. The tilted exposure ends up just being anywhere on the square.

I often get the message:

libCV failed: redoing tilt 0.45
Bad libCV result: too much rotation

It then repeats the correlation...But in the end, it ends up with a matrix with a huge shift!

Target out of bounds, so cannot check error
Moving...
handling modeled stage posoition request from RCT

I changed the min/max feature size so that the features found are in the hundreds. What I noted during test "acquisitions" in the RCT node is that the features found jump quite a bit.

What can I do to make this work? Does it make sense to upgrade to Leginon 2.0 to fix this?

Anchi, Jim, I have sent you a link to some hl images - one with which the correlation works and two with which it does not. What kind of correlation parameters would you recommend to make the latter two work (those are to independent samples, not tilt pairs)? I could also send you tilt pairs when I am at the microscope again, but maybe you already know by looking at the untilted ones.

Thanks so much for your help
Matthias


Replies (1)

Re: RCT problems - Added by Anchi Cheng over 14 years ago

Hi, Matthias,

First, upgrade to 2.0 probably will not fix the problem. No improvement was made on the application you use there.

I looked at the images you sent. Just like you've described, your test grid is a grating which provides plenty of strong feature for libcv's feature finding algorithm to get it correctly. Your real grid, being a negative stain on continuous carbon, it is quite smooth and not nearly as easy for the algorithm. Even though there are some clear dirt, that is not a lot considering that the algorithm attempts to find several hundreds. I think that is the reason for the failure. You might want to try a different magnification where you can see more feature. The sq, hl mags we suggested in the manual are just suggestions. For rct acquisition which uses stage movement to get to the target, you don't have to limit yourself to 5000x, you can go to lower magnification at that step. The other extreme of the solution is to make the algorithm to look for the larger features so that it concentrates on the obvious. This will give fewer features but might work for you.

The other suggestion is to add more stable features to your grid. I don't mean put your grid in a sandbox to dirty it up <img src="{SMILIES_PATH}/icon_biggrin.gif" alt=":D" title="Very Happy" /> . You can try using a layer of of carbon on C-flat or quantifoil grid and then shoot images in the hole where there is only one layer of carbon.

Anchi

    (1-1/1)