SCIInstitute
Intensity Sampling for Data Augmentation
Add an optional step to data augmentation that uses histogram matching to create augmented images with different intensity profiles.
We tried two different method, one with adding pixel wise to the intensity data before doing the shape augmentation. Second we warped each shape to mean shape and then sample from the intensity distribution (with the PCA distribution), finally transform again to the shape+ intensity space.
What needs to be done is histogram matching of the generated images to random images from the distribution.
This is low priority and I haven't been working in it. @sheryjoe is this related to what anyone else is working on? Can we reassign it?
@j-void we could consider intensity-based data augmentation for DeepSSM and your work.
@sheryjoe Is this issue still relevant? Should we clarify it based on @j-void and @mahimoksha's adversarial work?
@jadie1 this is still relevant and can be useful when we have heterogenous intenstiy profiles (e.g. LGE), you can assign @mahimoksha to this as this is related to his augmentation research.