We use a custom branch of batchgeneratorsv2 for data augmentation in our training pipeline. We are looking for additional augmentations that may help with scroll specific data. Implement one (or more) of these, and conduct an ablation study on its performance. Augmentations ideally should be on gpu and speed is important.
Some ideas:
Decohesion:
- in dense areas the beam is scattered, producing an image that looks "blurred" or "smeared" from previous layers
Warping:
- The scroll in some areas is heavily warped. Simple grid deformations / elastic deformations do not tend to represent the type of warping present on scroll data. Implement an augmentation that can take a relatively straight chunk and augment it such that it appears similar to a more warped region
Squeezing/Pulling:
- Implement an augmentation that can create accurate compressed regions from normal data
Any additional augmentations that may improve performance are welcome
We use a custom branch of batchgeneratorsv2 for data augmentation in our training pipeline. We are looking for additional augmentations that may help with scroll specific data. Implement one (or more) of these, and conduct an ablation study on its performance. Augmentations ideally should be on gpu and speed is important.
Some ideas:
Decohesion:
Warping:
Squeezing/Pulling:
Any additional augmentations that may improve performance are welcome