Random cropping with scaling option for super-resolution data augmentation
See original GitHub issueIs your feature request related to a problem? Please describe. In super-resolution networks, a low resolution input is upsampled using a neural networks often with an integer factor (e.g. x2, x3, x4 etc). When performing data augmentation, a good approach is to use random cropping with fixed size, especially if input images are bigger than what can fit in memory for network activations.
Describe the solution you’d like To address the data augmentation issue of random cropping, a possible solution could be to get a random crop window in low resolution image and apply the same (but scaled up) window to target ground truth image.
For example, in a x4 upsampling network cropping a 56 x 56
window in low res input will correspond to 224 x 224
window cropping in target ground truth
Describe alternatives you’ve considered Writing my own MONAI transforms, or manually cropping data
Additional context This may be useful for super-resolution, upsampling or demosaicing networks that typically take low res input and up resolve it
Apologies if this is addressed by an existing transform, in which case can someone guide me on how I can achieve the above with an existing transform…
Issue Analytics
- State:
- Created a year ago
- Reactions:1
- Comments:5 (5 by maintainers)
Top GitHub Comments
not at the moment @masadcv. but as we have the MetaTensor implementation on dev, I hope it could be a universal solution of specifying roi_sizes in terms of physical units instead of number of voxels (for example the
hires_scale
should be computed from MetaTensor’spixdim
property). would you still be interested/have the bandwidth to contribute?Hi @wyli , Just wanted to check if anyone is working on this? If not I will like to contribute with this transform…