In recent DL work, consistency regularization consists on ensuring the same output while applying transformations on the input (eg: rotation, translation). Consistency regularization can notably be done with a student-teacher approach.
I'm wondering if we could possibly get inspiration from this approach for our contrast-agnostic segmentation? Eg: further alter the image contrast with non-linear signal scaling (randomize the target shape of the image histogram)?
In recent DL work, consistency regularization consists on ensuring the same output while applying transformations on the input (eg: rotation, translation). Consistency regularization can notably be done with a student-teacher approach.
I'm wondering if we could possibly get inspiration from this approach for our contrast-agnostic segmentation? Eg: further alter the image contrast with non-linear signal scaling (randomize the target shape of the image histogram)?