( 참고 : 패스트 캠퍼스 , 한번에 끝내는 컴퓨터비전 초격차 패키지 )

Data Augmentation (2)

Categories of Data Augmentation

  • (1) Rule-based
  • (2) GAN-based
  • (3) AutoML-based


(2) GAN-based

for more natural data augmentation

  • ex) medical image augmentation
  • ex) SinGAN : Single Image GAN
  • ex) GANGealing : GAN-Supervised Alignment


a) Medical Image Augmentation

GAN-based Synthetic Medical Image Augmentation for increased CNN Performance in Liver Lesion Classification, Frid-Adar et al., NIPS 2017

  • evaluate based on classification ( 3-class )

figure2


b) SinGAN

SinGAN : Learning a Generative Model from a Single Natural Image, Shaham et al., ICCV 2019, best paper award

  • UNconditional GAN with single Image

  • use multi-scale pyramid & Fully-convolutional GANs

    • GLOBAL & LOCAL information with pyramid

    figure2

  • manipulation task

    figure2

  • result :

    figure2

  • generator : small capacity ( to not memorize the training data )
    • ( 3x3 conv - BN - LeakyReLU ) x 5
  • Loss Function : Adversarial Loss + Reconstruction Loss

figure2


c) GANGealing

GAN-Supervised Dense Visual Alignment, Peebles et al., CVPR 2022

  • Supervision with images generated by GAN
  • architecture :
    • (1) pre-trained GAN ( LEFT )
    • (2) pre-trained conditional GAN ( RIGHT )
    • (3) STN (Spatial Transformer) \(\rightarrow\) update this one!
  • Step 1) generate images with (1) GAN
  • Step 2) generate images with (2) GAN, given certain condition \(c\)
  • Step 3) Align images from Step 1), using STN
  • Step 4) Loss ( images from step2, images from step 3 )

figure2


Cons :

  • only applicable in class-specific GAN
  • unapplicable when high pose variation

Categories:

Updated: