Data Augmentation Techniques

Transforming the data in various ways. Generally, if human can still understand the meaning then the transformation is okay. Can add various transformations together.

• Mixup
• Progressive resizing
• Image flipping (but maybe not vertical)
• random crop. CutMix: mixing/cutting different poritions together
• Cowmix: using "cow" mask to add occlusions (occlusions can be from different image or random pixels. If from other labeled image, ground truth can be proportional from both images
• CutOut: removing random square
• Color jitter: Adding/subtracting random values to RGB channels
• Geometric transformations: translation, rotation, shear, scale
• FixMatch: Use labeled, get weakly augmented to get prediction, then if prediction is above threshold use it as ground truth for strongly augmented. Total loss has two terms: 1. The cross entropy loss of the labeled example and it's weakly augmented prediction, and 2. The cross entropy loss between the hard pseudo-label of a weakly augmented image above some threshold and the prediction of the strongly augmented image.

blah