Multiview Reconstruction Plugin Help

[nellypledge]

Hi guys, I hope the data treats you well.

I've been playing around with some SPIM data and the multiview reconstruction plugin.  I've been having trouble registering some data sets from some cleared tissue data sets that fill the field of view, and hence I have no registration beads to work with.

I've been trying to use the redundant geometric local descriptor matching option for the type of registration in the basic registration options, which I assume is the correct option (I recall hearing Stephan mention it in his tutorial presentation).  On the ImageJ.net help page I see that it requires the info on the rotations.  It knows the angles, and so I'm not sure if it already applies the rotations for me...?  On using the apply transformations tool to rotation each data set the required amount around the x-axis I see in the BigDataViewer that it rotates on the corner of the volume.  Does the translation-invariance mean that the volumes can be translated but not rotated?  I also tried around the y-axis in case I had them mixed up but I saw something equally 'star shaped' in the BigDataViewer.  Any tips on the manual transformation controls would be very helpful.

In the following windows there are many options which I am equally unsure about.  Any info on the nitty gritty of all the options and parameters would be much appreciated (Fix Tiles, Map back, Transform model, Regularization model and settings, and then the regularization settings too).  Alternatively, if it's an option, if anybody has any documents and/or presentations on the related topics to help bring me up to speed that would be great.  I've started to read some


Thanks in advance for your time.

Neil

________________________________

This e-mail message (including any attachments) is for the sole use of
the intended recipient(s) and may contain confidential and privileged
information. If the reader of this message is not the intended
recipient, you are hereby notified that any dissemination, distribution
or copying of this message (including any attachments) is strictly
prohibited.

If you have received this message in error, please contact
the sender by reply e-mail message and destroy all copies of the
original message (including attachments).

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html

Hi Neil,

I gave a short tutorial at Janelia this week which was recorded and I will try to make it accessible as soon as possible. In the meantime, I am trying to answer some of your questions.

How do you import the data into the Multiview Reconstruction? If it is via loading the CZI dataset, then, yes, it should apply the approximate rotation already. It does not really matter around which point it rotates as the matching is done translation-invariant. However, for other datasets (e.g. TIFFs) one needs to figure out the right rotation (x-axis or y-axis, e.g. +90 or -90 degrees), which depends on the microscope itself. I recently added a simple center of mass alignment which might help as after applying a transformation it can more or less overlap the views. I realise that there should maybe be a tool that helps figuring this out using Max-Intensity projections in all potentially correct orientations. It is on the to-do list now.

The following parameters mostly are a tradeoff between the likelihood of the transformation identified being correct and identifying a transformation at all:

Fix Tiles … Usually one tile/view/stack is not modified during global optimization. My suggestions is to not change this if you do not understand the details of the global optimization procedure.
Map Back … Only matters if no tile/view/stack is fixed

Transformation Model - Affine/Rigid/Translation … well, that depends on the task. Usually Affine, regularised with Rigid makes most sense
Number of Neighbours … More neighbours gives more reliable, but less matching descriptors (check out the video of my presentation on the principle: https://www.youtube.com/watch?v=IupXS_On2rg)
Redundancy … Tolerante potential misdetected blobs - more redundancy will find more potential matches, but is slower and might find more that are not correct
Significance (N-times) … For each descriptor we find the best matching and second best matching one in the other view. We consider it a potential match if the best one is N-times better than the second best one. So a higher number results in more reliable matches, a lower number in more matches - play with this one first if you identify no transformation model
Allowed RANSAC error … All potentially corresponding descriptors of two views are subject to a (transformation in this case) model-based outlier removal, where this is the maximally allowed error so it still is considered as “agreeing” on the same model. Keep in mind that sometimes the z-resolution is way worse than xy - this pixel error is isotropic coordinates, so usually in xy pixel units.

Cheers,
Stephan

---

Dr. Stephan Preibisch
Group Leader

Berlin Institute of Medical Systems Biology of the MDC
Building 89, 1.08b

email: [hidden email]<mailto:[hidden email]>
web: http://www.preibisch.net/<http://fly.mpi-cbg.de/preibisch>

On 30 Jun 2016, at 20:11, Anthony, Neil <[hidden email]<mailto:[hidden email]>> wrote:

Hi guys, I hope the data treats you well.

I've been playing around with some SPIM data and the multiview reconstruction plugin.  I've been having trouble registering some data sets from some cleared tissue data sets that fill the field of view, and hence I have no registration beads to work with.

I've been trying to use the redundant geometric local descriptor matching option for the type of registration in the basic registration options, which I assume is the correct option (I recall hearing Stephan mention it in his tutorial presentation).  On the ImageJ.net<http://imagej.net> help page I see that it requires the info on the rotations.  It knows the angles, and so I'm not sure if it already applies the rotations for me...?  On using the apply transformations tool to rotation each data set the required amount around the x-axis I see in the BigDataViewer that it rotates on the corner of the volume.  Does the translation-invariance mean that the volumes can be translated but not rotated?  I also tried around the y-axis in case I had them mixed up but I saw something equally 'star shaped' in the BigDataViewer.  Any tips on the manual transformation controls would be very helpful.

In the following windows there are many options which I am equally unsure about.  Any info on the nitty gritty of all the options and parameters would be much appreciated (Fix Tiles, Map back, Transform model, Regularization model and settings, and then the regularization settings too).  Alternatively, if it's an option, if anybody has any documents and/or presentations on the related topics to help bring me up to speed that would be great.  I've started to read some


Thanks in advance for your time.

Neil

________________________________

This e-mail message (including any attachments) is for the sole use of
the intended recipient(s) and may contain confidential and privileged
information. If the reader of this message is not the intended
recipient, you are hereby notified that any dissemination, distribution
or copying of this message (including any attachments) is strictly
prohibited.

If you have received this message in error, please contact
the sender by reply e-mail message and destroy all copies of the
original message (including attachments).

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html


--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html

[TimFeinstein]

Thank you Stephan!  I had questions similar to Neil and eagerly look
forward to seeing your tutorial.

All the best,


Tim

Timothy Feinstein, Ph.D.
Research Scientist
University of Pittsburgh Department of Developmental Biology





On 7/1/16, 9:18 AM, "ImageJ Interest Group on behalf of
[hidden email]" <[hidden email] on behalf of
[hidden email]> wrote:

>Hi Neil,
>
>I gave a short tutorial at Janelia this week which was recorded and I
>will try to make it accessible as soon as possible. In the meantime, I am
>trying to answer some of your questions.
>
>How do you import the data into the Multiview Reconstruction? If it is
>via loading the CZI dataset, then, yes, it should apply the approximate
>rotation already. It does not really matter around which point it rotates
>as the matching is done translation-invariant. However, for other
>datasets (e.g. TIFFs) one needs to figure out the right rotation (x-axis
>or y-axis, e.g. +90 or -90 degrees), which depends on the microscope
>itself. I recently added a simple center of mass alignment which might
>help as after applying a transformation it can more or less overlap the
>views. I realise that there should maybe be a tool that helps figuring
>this out using Max-Intensity projections in all potentially correct
>orientations. It is on the to-do list now.
>
>The following parameters mostly are a tradeoff between the likelihood of
>the transformation identified being correct and identifying a
>transformation at all:
>
>Fix Tiles Š Usually one tile/view/stack is not modified during global
>optimization. My suggestions is to not change this if you do not
>understand the details of the global optimization procedure.
>Map Back Š Only matters if no tile/view/stack is fixed
>
>Transformation Model - Affine/Rigid/Translation Š well, that depends on
>the task. Usually Affine, regularised with Rigid makes most sense
>Number of Neighbours Š More neighbours gives more reliable, but less
>matching descriptors (check out the video of my presentation on the
>principle:
>https://na01.safelinks.protection.outlook.com/?url=https%3a%2f%2fwww.youtu
>be.com%2fwatch%3fv%3dIupXS_On2rg&data=01%7c01%7ctnf8%40pitt.edu%7c4ee6e23f
>893a4cbe6b4808d3a1b861a1%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=WB5m
>Ta3CVM%2fz7MvJHzxGY5bwBgAJumx4tbZaikdFTT8%3d)
>Redundancy Š Tolerante potential misdetected blobs - more redundancy will
>find more potential matches, but is slower and might find more that are
>not correct
>Significance (N-times) Š For each descriptor we find the best matching
>and second best matching one in the other view. We consider it a
>potential match if the best one is N-times better than the second best
>one. So a higher number results in more reliable matches, a lower number
>in more matches - play with this one first if you identify no
>transformation model
>Allowed RANSAC error Š All potentially corresponding descriptors of two
>views are subject to a (transformation in this case) model-based outlier
>removal, where this is the maximally allowed error so it still is
>considered as ³agreeing² on the same model. Keep in mind that sometimes
>the z-resolution is way worse than xy - this pixel error is isotropic
>coordinates, so usually in xy pixel units.
>
>Cheers,
>Stephan
>
>---
>
>Dr. Stephan Preibisch
>Group Leader
>
>Berlin Institute of Medical Systems Biology of the MDC
>Building 89, 1.08b
>
>email:
>[hidden email]<mailto:[hidden email]>
>web:
>https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fwww.preibi
>sch.net%2f&data=01%7c01%7ctnf8%40pitt.edu%7c4ee6e23f893a4cbe6b4808d3a1b861
>a1%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=ZkEz%2fvNGbgC3Ae9zU0hL5mV5
>dDYCCk0WQNIpriJ5HFY%3d<https://na01.safelinks.protection.outlook.com/?url=
>http%3a%2f%2ffly.mpi-cbg.de%2fpreibisch&data=01%7c01%7ctnf8%40pitt.edu%7c4
>ee6e23f893a4cbe6b4808d3a1b861a1%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sda
>ta=X9JW9mZxTd2ALQJEMq5XXjuZE5DqX6GlIUVVCR%2brAmQ%3d>
>
>On 30 Jun 2016, at 20:11, Anthony, Neil
><[hidden email]<mailto:[hidden email]>> wrote:
>
>Hi guys, I hope the data treats you well.
>
>I've been playing around with some SPIM data and the multiview
>reconstruction plugin.  I've been having trouble registering some data
>sets from some cleared tissue data sets that fill the field of view, and
>hence I have no registration beads to work with.
>
>I've been trying to use the redundant geometric local descriptor matching
>option for the type of registration in the basic registration options,
>which I assume is the correct option (I recall hearing Stephan mention it
>in his tutorial presentation).  On the
>ImageJ.net<<a href="https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2">https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2
>fimagej.net&data=01%7c01%7ctnf8%40pitt.edu%7c4ee6e23f893a4cbe6b4808d3a1b86
>1a1%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=3xjyu%2b1DYZu9fgXM8WHQAsx
>b8RQrl%2bGm%2b3ZLkncueXU%3d> help page I see that it requires the info on
>the rotations.  It knows the angles, and so I'm not sure if it already
>applies the rotations for me...?  On using the apply transformations tool
>to rotation each data set the required amount around the x-axis I see in
>the BigDataViewer that it rotates on the corner of the volume.  Does the
>translation-invariance mean that the volumes can be translated but not
>rotated?  I also tried around the y-axis in case I had them mixed up but
>I saw something equally 'star shaped' in the BigDataViewer.  Any tips on
>the manual transformation controls would be very helpful.
>
>In the following windows there are many options which I am equally unsure
>about.  Any info on the nitty gritty of all the options and parameters
>would be much appreciated (Fix Tiles, Map back, Transform model,
>Regularization model and settings, and then the regularization settings
>too).  Alternatively, if it's an option, if anybody has any documents
>and/or presentations on the related topics to help bring me up to speed
>that would be great.  I've started to read some
>
>
>Thanks in advance for your time.
>
>Neil
>
>________________________________
>
>This e-mail message (including any attachments) is for the sole use of
>the intended recipient(s) and may contain confidential and privileged
>information. If the reader of this message is not the intended
>recipient, you are hereby notified that any dissemination, distribution
>or copying of this message (including any attachments) is strictly
>prohibited.
>
>If you have received this message in error, please contact
>the sender by reply e-mail message and destroy all copies of the
>original message (including attachments).
>
>--
>ImageJ mailing list:
>https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fimagej.nih
>.gov%2fij%2flist.html&data=01%7c01%7ctnf8%40pitt.edu%7c4ee6e23f893a4cbe6b4
>808d3a1b861a1%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=lY26Yp%2bxw4ssu
>CSrl76s43JNBfb11cvkgSVx%2bcGGdn4%3d
>
>
>--
>ImageJ mailing list:
>https://na01.safelinks.protection.outlook.com/?url=http%3a%2f%2fimagej.nih
>.gov%2fij%2flist.html&data=01%7c01%7ctnf8%40pitt.edu%7c4ee6e23f893a4cbe6b4
>808d3a1b861a1%7c9ef9f489e0a04eeb87cc3a526112fd0d%7c1&sdata=lY26Yp%2bxw4ssu
>CSrl76s43JNBfb11cvkgSVx%2bcGGdn4%3d

--
ImageJ mailing list: http://imagej.nih.gov/ij/list.html

[nellypledge]

Hi Stephan, thanks for the detailed message; I really appreciate your time.

My files are tifs taken rotated around the y axis (I believe my xyz are from the single detection lens perspective; I do have two cameras, and the image may be flipped, but I assume that doesn't affect things).

When I use beads around a sample I don't need to rotate the image (I assume that because I use the 3D geometric hashing option with takes care of the rotation?), and it finds things well.  If I have no reference beads, and select redundant geometric local descriptor matching, can I confirm that I need to go to apply transformations and make things overlap to some degree?

...  interlude for some testing clicky clicky  ...

With some time I have managed to align three of four views (see below Aside), but have lost ability to bring the forth into the right place.  Either way, with some work I can register the different views to some degree, which is great!

I will persevere.  Any manual transformation tips much appreciated.

Also, I'll look forward to the latest tutorial presentation.

Thanks
Neil


[Aside:  Do you have any tips on how to handle things in the BigDataViewer?  My manual transformations are instantly ruined if I click the left button, and I'm loosing my mind trying to figure out how far to step things (together with being able to see more than one view but move only one view).  Is there an option for inactive but visible?  I've tried groups but everything is B&W again and colour really helps.]