Data sets

Eliceiri's data set

Images are at MIDA1 server under: /home/elisabeth/HighRes_Splits.zip

The folder contains a few folders, and it would be great if you could do the following:

• "CNN_Training" (40 images)
• "CNN_Training_Crops" (40 images)
• "CNN_Validation" for any hyper parameter tuning we need for the proposed method or the GAN training. (25 images)
• "Registration_Tuning" for any hyper parameter tuning of registration methods (7 images)

In "CNN_Training" the images are 2048x2048, each a pair of brightfield and SHG, they would belong by naming like for example "1B_A2_BF.tif" and "1B_A2_SHG.tif", so the first part is the image name and the last gives you which modality - BF or SHG. They are all perfectly registered and of the same size. "CNN_Training_Crops" are 834x834 center crops of those images that we decided to all train on.

For final evaluation we have the folder "Registration_UltimateTesting" and in there the following folders:

• "Both_Transformed_and_References_Patches" (134 images)
• "RotationOnly_Transformed_and_References_Patches" (134 images)
• "TransformationOnly_Transformed_and_References_Patches" (134 images)

Here the naming differs a bit, as that at the end of the filename there's either a "_R" or a "_T" for reference or transformed image. Again there's two of each modality. These images are only 834 × 834 to avoid any border effects due to the transformations.

The transformations performed on the images:

• "RotationOnly_Transformed_and_References_Patches": random rotations in [-60,+60] degrees
• "TransformationOnly_Transformed_and_References_Patches": random translations in [-614,614] pixels (~=[-30,30]% of original image size)
• "Both_Transformed_and_References_Patches": both such rotations and translations

For all three, the transformations are picked in random, but such that a third of them qualifies as "small transformations", a third as "medium" and a third as "large".

• "Small" is defined by a transformation that results in an average displacement of corners <100px
• "Medium" is defined by a transformation that results in an average displacement of corners in [100px, 200px]
• "Large" is defined by a transformation that results in an average displacement of corners >200px

Zurich data set