The original repository uses conda for environment management, the environment does work well with CUDA 12. Also editable installs do not work because torch can not be found.
TODO: Find a better way for editable installs
Workaround
Use python setup.py install inside every every submodule package, include the missing header files (look here)
TODO: create a shell script that does all the installation or automate it someway
Densification Threshold
0.044
0.0002
NOTE: DICOM to RGB https://medium.com/analytics-vidhya/how-to-convert-grayscale-dicom-file-to-rgb-dicom-file-with-python-df86ac055bd
How can we do that ?
Current Errors
Currently we are facing an error about accessing illegal memory in CUDA, I’ll try to transform the DICOM images to RGB and let’s see if this will fix the issue.
An RGB conversion is to convert to the interval [-1000, 2000] then do normalize between 0 and 1
Questions
How can I get projections and directions from the 3D volume, I am getting the 3D volume by stacking all the slices.
Also can I apply frustum culling to do segmentation or sime kind of stuff like that or introduce a new parameter.
I introduce a depth parameter and optimize the view by giving the current depth, but achieve this I’ll have to play with rasterizer.