Re: Elastic layer alignment in TrakEM2 - problem with correspondences

[Klaus Griebel]

Hello again,

I've managed to get some testing with the Test Block Matching Plugin done, but I'm ca. 120 tries in and so far haven't accomplished much.  I've exported a subset of my series as a RBG green stack (0,255,0) and scaled it to 0.05 (as the ratio of pixels in x,y to z is about 1:20 -> 0.05), which I hope I understood correctly.

I have attached the results and according settings that I interpret to be the best so far. The reason for this is, that all other tries resulted in pictures with either a lot more white gaps or a lot less smooth fields. Reducing the block and search size yielded smaller fields to random color pixels while increasing block and search size yielded less and less hits in the periphery.

Unfortunately the only correlation filter parameter that helped improve the result (according to my untrained jurisdiction) was reducing the min PMCC r. So far -0.30 appears to be the best fit, I've only tried reducing it to -0.50 once but more white gaps are showing.

My problems are mainly that changing parameters according to the explanation from the Wiki page other "negative" occurences arise immediatley i.e. reducing white gaps leads to more false positives in terms of saturated patches or strengthening filters to reduce saturated colors generates more white patches.

As I am not capable of progressing on my own I'd like to ask for help again.

Greetings,
Klaus


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

[Saalfeld, Stephan]

Hi Klaus,

sorry for being silent.  Pretty busy days right now.  Can you share some
images?  First glance guessing, all of that may be wrong because I do
not see the data:

1. Your montages are bad.  I can see the tile boundaries as jumps in the
block matching results.

2. The search radius is too small, color saturated, flowing over where
displacements are larger.

3. Mesh resolution is way too high.  Your are basically killing time.
If the deformation is smooth elastic you need ~4x less mesh resolution,
but I guess your montages are bad.

If 1. is true, it will not work.

For everything else, I need to see the data.

Best,
Stephan



On Wed, 2015-07-29 at 04:15 -0400, Klaus Griebel wrote:

> Hello again,
>
> I've managed to get some testing with the Test Block Matching Plugin done, but I'm ca. 120 tries in and so far haven't accomplished much.  I've exported a subset of my series as a RBG green stack (0,255,0) and scaled it to 0.05 (as the ratio of pixels in x,y to z is about 1:20 -> 0.05), which I hope I understood correctly.
>
> I have attached the results and according settings that I interpret to be the best so far. The reason for this is, that all other tries resulted in pictures with either a lot more white gaps or a lot less smooth fields. Reducing the block and search size yielded smaller fields to random color pixels while increasing block and search size yielded less and less hits in the periphery.
>
> Unfortunately the only correlation filter parameter that helped improve the result (according to my untrained jurisdiction) was reducing the min PMCC r. So far -0.30 appears to be the best fit, I've only tried reducing it to -0.50 once but more white gaps are showing.
>
> My problems are mainly that changing parameters according to the explanation from the Wiki page other "negative" occurences arise immediatley i.e. reducing white gaps leads to more false positives in terms of saturated patches or strengthening filters to reduce saturated colors generates more white patches.
>
> As I am not capable of progressing on my own I'd like to ask for help again.
>
> Greetings,
> Klaus
>
>
> --
> ImageJ mailing list: http://imagej.nih.gov/ij/list.html

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

[Klaus Griebel]

Hi Stephan,

thank you very much for your answer. I have attached three arbitrarily chosen layer exports that i exported as .jpeg at 20% scale (adjusting the filesize to post them here).

Regarding point 1) Yes and No... I think.

Considering the alignment not all of the ones shown in the Block Matching results were bad. I only later in the process found the "perfect" parameters for elastic montaging, so from layers ca. 70-130 the montages overlap perfectly. Unfortunately I have not gone back for every layer yet, therefor not all are perfectly aligned intra layer.

Considering the image borders there is one issue I think you can see from all example images. Most of my images have a brightness gradient that stems from the TEM we are using (mostly very old components) and even post processing every image in Photoshop could not undo this easily.

Furthermore the blending function of TrackEM2 does not work with my images as I have to apply coordinate transforms to my images prior to blending and my version would not let me (Error message states that this function will be added in the future). Applying it before coordinate transformation leads to the merging of all the images in the layer to one big image while the distorted overlaps are not corrected.

On a side note: I am using TrackEM2 on a Mac and my Fiji Updater seems to be malfunctioning... Is blending with coordinate transformed images implemented in the latest version?

Regarding points 2) and 3)

I will try reducing the resolution from my 128 to 32 and increase the search size in the Test Plugin.

Thank you very much.

Kind regards,
Klaus

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