If you don’t have any preferred root branch then just choose any larger branch that is connected to most others. Start with a cropped model and apply decimation to have no more than a few ten thousand points. You can then experiment with gradually increasing the region of interest size and tune decimation parameters to reach a good tradeoff between quality and amount of data you get, and computation time on your system.
There may be tricks that long-time VMTK users know about how to handle large networks, so please reach out to them and then report back here that you learned.
I couldn’t exactly understand how to obtain a cropped model. In the segment editor I try to use “Scissors” , select operation and shape (circle) required. How should I crop the volume after this step? In one of my previous post, you had mentioned Scissors will only blank the voxles. So I am not sure how to crop.
Could you please briefly explain how decimation works? I find a Reductiontab that shows (0.8) by default. Does this mean 0.8 percent of the original number of points are removed?
May I know which feature can be used to get an estimate of the number of points?
Definitely, I have reached out to them a week back and wrote to them for the second time today. I will wait for their response and share my learnings here.
If you crop to reduce memory usage then you need to crop the input volume (before you start segmentation) using Crop volume module. If you crop so that you have a network that VMTK can process faster, you can further cut off parts of the image using Scissors tool.
Reduction refers to the requested fraction of points to be removed. 0.8 means that you request removal of 80% of points. Probably you want to use values in the 0.9-0.99 range.
Number of points are shown in the tooltip when you hover over your model in Data module. Some more information is shown in Models module / Information section.
Explain them in 2-3 sentences who you are, what your project is about, and why it is important. Also attach a screenshot link to example labelmap and segmented model files. Then you have a better chance to get an answer. You may also keep asking 1-2x a week for a couple of weeks, telling what you have tried, and how well those worked.