AugMix

  • @hendrycksAugMixSimpleData2020
  • using the input image itself
  • t transforms (translate, shear, rotate and etc) the input image and mixes it with the original image
  • Image transformation involves series of randomly selected augmentation operations applied with three parallel augmentation chains.
  • Each chain has a composition of operations that could involve applying, for example, translation on input image followed by shear and so on
  • The output of these three chains is three images mixed to form a new image.
  • This new image is later mixed with the original image to generate the final augmented output image,
  • while we considered mixing by alpha compositing, we chose to use elementwise convex combinations for simplicity. The k-dimensional vector of convex coefficients is randomly sampled from a Dirichlet(α, … , α) distribution.
  • Once these images are mixed, we use a “skip connection” to combine the result of the augmentation chain and the original image through a second random convex combination sampled from a Beta(α, α) distribution. The final image incorporates several sources of randomness from the choice of operations, the severity of these operations, the lengths of the augmentation chains, and the mixing weights
  • Jensen Shannon Divergence Consistency Loss