search

wustl-cig / DOLCE

DOLCE, ICCV2023. Pytorch Implementation.
MIT License
28 stars 5 forks source link

Questions About Test Data Generation code and 150-Degree Results #7

Closed uigeunahn closed 3 days ago

uigeunahn commented 1 month ago

Hello,

I have a few questions regarding the test data provided:

Test Data Generation: Could you please provide the code for generating the test data, specifically the FBP and RLS images? Having access to this code would be extremely helpful for my analysis.

150-Degree Data: I noticed that data for the 150-degree is provided, so I ran the model with this data. It seems like the window level might have changed compared to other result of 60-degree, 90-degree, 120-degree as it seems little bit darker. Additionally, the performance for the 150-degree setup is similar to or slightly worse than the 120-degree configuration. Could you provide any insights or reasons why this might be happening? Any suggestions for improving performance with not only 150-degree data but also less (ex 45-degree or 30-degree) degree would also be appreciated.

Thank you for your work! I really learned a lot from your paper.

hkimdavis commented 3 days ago

The code to generate the test data used FBP and RLS functions in Livermore Tomography Tools (LTT) which requires a license. We released a python software package called LEAP as an open source, which provides forward projection, FBP, RLS, etc. You can find some examples of FBP and RLS using LEAP: GitHub.com/LLNL/LEAP.

JiamingLiu-Jeremy commented 3 days ago

Regarding the second question, there is no simple answer about the performance difference between 120 and 150 degrees, as it depends on the object configuration and the object shape/orientation with missing projection angles due to the limited angle setup. One potential limitation of the experiments is the fixed angular range, meaning the starting angle is always 0 degrees, resulting in an angular range of 0-150 degrees. Additionally, extreme LA settings are still a challenging issue, and we expect new conditional training of DPMs and sampling algorithms to help. We are working towards this direction.