Integrate LabKit in TrackMate to edit segmentation results #302
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
And remove Gson from now. Apparently there are incompatibilities from the (same) declaration in LabKit pom, which triggers the errors I complained about today on the forum.
This launcher catches what time-point is currently diplayed, and wrap all information needed for LabKit to edit the spot it contains, as labels. While the user edits the labels in LabKit, the TrackMate UI is disabled. Once the user closes the LabKit window, the modifications they made are inspected. This class compares the new labels with the previous ones, and can determine whether a spot has been added, removed or modified. In the last case it updates the model with the modified spots, and reintroduces it in the tracks as it should. All features of modified spots and their edges are recomputed. If a label has several connected components, they are added as separate spots. The one closest to the original spot is reintroduced in the tracks. The label names are important: they are used to retrieve the original spot id and the original spot shape for comparison. If the user modifies a label, it will be perceived as a new spot instead of a modified one. Tested in 2D so far.
…ard. But because the other table, trackscheme and bvv buttons take too much place, we don't see it without resizing the window.
Otherwise the button for the editor is not visible without resizing the window. I also had to programmatically resize the main TrackMate frame after display so that components are properly aligned. Such a hack for something simple... Also give the editor button a proper name and a temporary icon.
- We don't depend on labels anymore, but directly operate and compare the index images (before modification and after). Because the index is directly related to the spot ID, we can get a match from previous spot to novel spot in an easy manner. - The spots from the edited version are created directly from the novel index image, using something adapted from the label image detector code, so again, just one pass. We use the fact that we can provide it with a 'quality' image, and read the index of the label image 'under' the spot and write it into its quality value. This way we can retrieve the id of the matching previous spot in an easy manner. - The price to pay for not working with labels anymore is that we don't have access to the label name, but that's life. - We make only one pass over the image to collect the ids of the spots that have been modified, instead of one pass per spot. Also, this pass is multithreaded (thanks LoopBuilder). - I have also learned that I should not use weakListeners() if I am doing something with threads inside the listener. Using listeners() instead works, but I do not know why the other one does not. Probably something arcane with Java WeakReferences being collected. - As a result of all this the performance is much better than before and the 'return to TrackMate' should happen without the user noticing the process.
I wanted to use it to check whether the user has activated the 'LabKit' update site, but apparently the labkit jars are included with vanilla Fiji.
Let them choose to discard or commit the changes.
So that the image in the LabKit window opens with the same display settings than in the ImagePlus main view.
We don't need the Labeling wrapper.
The LabKit editor worked fine foe us but only in 2D. In 3D the labels imported from the spots into LabKit were off along the Z axis, as if the Z caibration was not handled properly. This was caused by the custom BDV showable I made missing some important preparation steps. I simply copied these steps from the working version of BDVShowabble in the core LabKit code.
As suggested by Matthias.
Shift-click on the button.
And remove debug code.
I am not so sure it is a good idea for the color...
and use the white LUT when the imp is not displayed as a Composite.
Noticed by @MiniMiette
Otherwise we get crashes when we have more than 4k labels. Which is not what Labkit is optimized for but we will see that in a second time. In case we change our minds on the backing integer type, right now the labkit launcher and importer classes are generic.
The re-importing of labels from Tabkit to TrackMate could fail for 2D images and labels with a large index. For instance, it failed consistently when trying to re-import labels with an index larger than 65643. This problem roots in the getSpots() method of LabkitImporter. It relies on a trick: We get the new label image, and create spots from this label image. But we want the new spots to keep track of the index in the label image they were generated from. For this, in 2D, we use the MaskUtils.fromLabelingWithRoi() method. These methods accept an image as last argument used to read a value in the label image within the spot, that is normally used for the quality value of the new spot. But the SpotRoiUtils.from2DLabelingWithRoi() method converted the extra image to ImagePlus (because I was lazy). So the label image was effectively cast on ushort for an IntegerType image, hence the problem with the max label being 65453. The solution is to rewrite the from2DLabelingWithRoi() so that it does not rely on converting to ImagePlus, but on good old iteration with imglib2.
Provided that the detector that is called is cancelable.
Make sure the detection preview panel is slanted at the bottom of its display.
The bug was causing weird issues with unedited spots being deleted, unedited spots being duplicated etc. It took me really long to understand the cause. It was hidden in the step where we go from a label image to a collection of spots. Because spots are polygons, with simplified contours, there might be some pixels on the edges of the object that are not strictly inside the label. In this importer, we read the label value in one go, by storing it in the QUALITY value of the spot, in the MaskUtils class. But since the spots have simplified contours, and since the QUALITY value is the maximal value iterated over, our approach might fail on border cases: - when the contout is approximated and include pixels from another object - and when this object has a label value higher than the lael of the spot. This commit include a temporary fix: we reiterate over the spot but takes the median value iterated over, to make sure we read the correct value for the label. Other attempts will follow, for reference. But a true fix involves making a method that returns a map from label value to spot.
This time we fix it by creating spots that do not have a simplified contours. In that case we strictly iterate over the pixels inside label and get the correct value. However the created spots have a pixelated aspect (contours are not simplified), which might not be what the user wants. We should let them choose. Still not the perfect solution, as mentionned in the previous commmit.
Use the new label immg to spot method to get spots AND the label they come from.
Also better message when closing the editor.
If a ROI exists in the input image when launching the LabKit editor, only the image and the spots present in the ROI are sent to the editor. Spots touching the borders are not imported.
Normally pressing space and dragging should move the view, as for normal ImageJ tools. The removed line was preventing it.
When editing a sub-portion of the image, the spots outside the ROI are not imported in the editor. It is possible that the user creates a new label in LabKit that will have the same id that an existing spot outside the ROI. Before this fix, the spots in that case were replaced by the new ones, creating a mess. The fix consists in reminding what spots are imported in the LabKit editor, then compariing to this list the new ids when reimporting from the editor.
Simplified, removing the menu, the segmentation features and the segmenter features. Somply done by copy-pasting Matthias classes and removing the lines with the features to prune.
This time it was due to confusion between a map from spot ID to spots, and a map from label value to spot. I fixed it by rewriting everything in terms of label values in the labeling. So we do not need to keep track of the spot ID anymore. The labels could be anything. They are still equal to spot ID + 1, because it is convenient to debug.
When editing the whole movie, if the label of a new spot was using the label of an existing spot in another time-point, the existing one was removed. Because the new spot was identified as a modification of the existing one. The solution is to pass to the importer only the list of existing spots in the current time-frame.
This could happen after loading a TrackMate file with specific display settings but no tracks.
Like the latest YOLO...
Mother class for CLI configurator that relies on an executable installed in a conda environment. This happens when the Python code we want to call does not have a Python module (that we could call with 'python -m') or is simply not a Python program. In that case: on Mac we resolve the env path, and append the executable name to build an absolute path to the executable. on Windows we simply activate the conda environment and call the executable assuming it is on the path.
A null key means that the argument should not appear in the settings map when the CLI will be deserialized to the map. This is useful e.g. for arguments of the CLI that must not be configured by the user, nor should not be serialized to disk. Such as the path where to save temp images.
0 is good enough as we require the quality to be a positive value.
|
Merged into #309 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.