Hi ultrack team,
firstly, thank you for developing such an amazing package! I have been implementing ultrack for my own data (early developing zebrafish embryo, 3D image volume from a laser scanning confocal microscope). Everything seems to be running fine, however I'm running into problems trying to tune the parameters for the tracking wrapper (i.e. segmentation & linking). I generated the input foreground and contour masks with the label to foreground conversion helper from stardist segmentation (which looks quite decent, even though the imaging depth is not the best), however when trying to track for more than 10 frames the tracking algorithm introduces an extremely high number of false positive divisions. It captures around 50% of my annotated groundtruth (which contains 19 divisions in a 10 frame subset, while introducing 4 false divsions). This unfortunately explodes when I go to longer tracking windows (~40 frames) to 150 false divsions on 46 gt annotated (division recall accuracy here increases to ~45%). I have tried adjusting the weight parameters but I haven't really gotten any better results than this. On top of this I need to set the divsion weight to something around -1e6 until I am actually not detecting any divsions anymore (this was more of a sanity check for myself to see what the ranges of this parameter might be for my data). Since this seems quite excessive, considering also your published config.toml where weights were much closer to the default values, I was wondering if I am missing something on my end, maybe an issue with the input data or something I can't think of.
I would greatly appreciate any input from your side or suggestions for adjustments I might try.
Please let me know if you want me to provide any data or more detailed descriptions!
Best,
Daniel
Hi ultrack team,
firstly, thank you for developing such an amazing package! I have been implementing ultrack for my own data (early developing zebrafish embryo, 3D image volume from a laser scanning confocal microscope). Everything seems to be running fine, however I'm running into problems trying to tune the parameters for the tracking wrapper (i.e. segmentation & linking). I generated the input foreground and contour masks with the label to foreground conversion helper from stardist segmentation (which looks quite decent, even though the imaging depth is not the best), however when trying to track for more than 10 frames the tracking algorithm introduces an extremely high number of false positive divisions. It captures around 50% of my annotated groundtruth (which contains 19 divisions in a 10 frame subset, while introducing 4 false divsions). This unfortunately explodes when I go to longer tracking windows (~40 frames) to 150 false divsions on 46 gt annotated (division recall accuracy here increases to ~45%). I have tried adjusting the weight parameters but I haven't really gotten any better results than this. On top of this I need to set the divsion weight to something around -1e6 until I am actually not detecting any divsions anymore (this was more of a sanity check for myself to see what the ranges of this parameter might be for my data). Since this seems quite excessive, considering also your published config.toml where weights were much closer to the default values, I was wondering if I am missing something on my end, maybe an issue with the input data or something I can't think of.
I would greatly appreciate any input from your side or suggestions for adjustments I might try.
Please let me know if you want me to provide any data or more detailed descriptions!
Best,
Daniel