New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Ellipsoid resulting model is bad #1082
Comments
It occurs to me that the ellipsoid dataset has changed. Can someone explain the difference between "v0" and "v1"? ShapeWorks 5.5 doesn't perform any better on the current ellipsoid dataset/parameters. @sheryjoe @jadie1 @iyerkrithika21 |
@akenmorris "v1" is the un-groomed ellipsoid datasets generated using the ShapeCohortGenPackage. It includes all the following modes of variations and option for generations:
Whereas the old dataset had only one mode of variation, I suppose. @jadie1, correct me if I am wrong regarding v0. |
@jadie1 We agreed before that we will provide two ellipsoid datasets, one aligned (similar to v0) and one not aligned (to demonstrate grooming). I think v1 is the latter. |
Yes ellipsoid-v1 is as @iyerkrithika21 described, so it should have more than one mode of variation. ellipsoid-v0 is the original data for which the ungroomed version is the same as the groomed. |
Why should v1 have more modes? The use case for this data should factor out modes related to misalignment.
This is becoming a bit confusing. Do we need to keep two aligned ellispoids data, one with one mode and the other with three modes? I think the aligned should only have one mode (pick one radius for variation) as this is supposed to be the simplest toy example for the optimization. |
Because the radius varies along all three axes there should be more modes of variation right?
I can remove ellipsoid_aligned-v1 or ellipsoid-v0. We don't have use cases that use both, I just left ellpsoid-v0 for users using older versions of ShapeWorks. This was the original request for creating ellpsoid_aligned-v1: "Can we regenerate v0 using the same code as v1 by disabling all transformations? This way we maintain consistency and have both segmentations and nice meshes to test both for fixed domains and cutting planes." I am not sure why the model has poor correspondence. I guess the parameters need to be retuned? Should I try with new ellipsoids that only vary along one radius? |
The grooming has problems as well: Notice ellipsoid 9 is not even centered with the others. Is this a problem in the ellipsoid example or in the Image library? @cchriste @archanasri @jadie1 @iyerkrithika21 |
The correspondence points don't look great for me either though, maybe it needs more smoothing? To me it looks like there are four modes of variation - the first is size, and the next three are radii in the x, y, and z direction. |
@jadie1, @iyerkrithika21, I just tried on a different Mac, fresh install of shape works 6.0 RC5, new conda installs, everything. Look at the alignment for the samples (no need to even look at DTs): ellipsoid.mp4 |
Oh, hah, @jadie1 , @iyerkrithika21 , you must turn off the |
Oh yeah I see now. I just ran the getting-started-with-grooming-segmentations.ipynb which uses the same data and it seems to me it has the same problem but not as bad. See the last video here: http://sciinstitute.github.io/ShapeWorks/notebooks/getting-started-with-grooming-segmentations.html (this video matches what I just got running it on the release branch). I guess we could go through the use case and notebook line by line and see where the grooming differs... |
Scratch that the notebook and use case distance transforms have the same center/alignment issue. The distance transforms from the notebook are just smoother. |
They are being aligned with ICP? |
Yes, the reference is ellipsoid_14 |
So do you think it ends up aligning ellipsoid 9 along only one side? I'm not sure how we should address this. After ICP, should apply a center of mass operation? @sheryjoe any thoughts? It seems wrong to capture this translation as part of the shape model. |
@akenmorris which translation? I can't see this in the video. Did you QC the center of mass alignmenet step? Also, number of ICP iterations could matter in some cases. |
The ellipsoids are not centered, so necessarily the shape model captures the translation. |
COM should center them. |
I agree, but is ICP de-centering them to match one side? |
The impact of ICP should be minimal wrt translation after COM, unless the reference is not really a median shape. |
What's the status here? Is anyone working on this? I think there are two separate issues, grooming and then optimization parameters. |
@jadie1 @iyerkrithika21 any update? |
@akenmorris when I try to optimize the ellipsoids in Studio without ICP, I get this message.
This error doesn't make sense because the |
@iyerkrithika21 , can you tell me more about how you're running these in studio? If I had to guess based on the filenames in the corners, you perhaps imported the binary segmentations, skipped grooming and clicked run? They need to be distance transforms. |
I imported the segmentations, ran groom in studio. and then clicked optimize. |
Will be fixed for release via issue #1133 |
@iyerkrithika21 , I've added an issue for what you saw: |
Results are terrible in 6.0-RC5. When did this happen?
6.0-RC5:
6.0Rc5-Ellipsoid.mp4
5.5:
5.5-Ellipsoid.mp4
Compare the variance chart:
6.0-RC5:
5.5:
The text was updated successfully, but these errors were encountered: