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

Data Augmentation (3)

Categories of Data Augmentation

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


(3) AutoML-based

  1. AutoAugment
  2. Population Based AutoAugment
  3. Fast AutoAugment
  4. Faster AutoAugment
  5. RandAugment
  6. UniformAugment
  7. TrivialAugment


1) AutoAugment

( AutoAugment : Learning Augmentation Policies from Data, Cubuk et al., CVPR 2019 )

  • like NAS (Neural Architecture Search), sample Augmentation Policy from RNN controller
  • Reinforcment Learning / Policy gradient
    • reward : validation accuracy

figure2

figure2

figure2


Cons : TOO MUCH COMPUTATION TIME

( \(\because\) policy gradient based on validation error )


2) Population Based AutoAugment

( Population Based Augmentation : Efficient Learning of Augmentation Policy Schedules, Ho et al., ICML 2019 )

Problem of AutoAugment : computationally infeasible

\(\rightarrow\) solution : Population Based AutoAugment

figure2


Characteristics

  • non-stationary augmentation policy schedules

    ( instead of fixed augmentaiton policy )

  • Exploration & exploitation

\(\rightarrow\) outputs an augmentation policy!

figure2

figure2


3) Fast AutoAugment

( Fast AutoAugment, Lim et al., NeurIPS 2019 )

Problem of AutoAugment : computationally infeasible

figure2


Solution : efficient search strategy, using density matching

  • concept of Bayesian Optimization ( Tree-structued Parzen Estimator (TPE) )

figure2


4) Faster AutoAugment

( Faster AutoAugment : Learning Augmentation Strategies using Backpropagation, Hataya et al., ECCV 2020 )

Motivation

  • make it differentiable ! DIFFERENTIABLE AutoAugment

figure2


Candidates of operations :

figure2


figure2

figure2


5) RandAugment

( RandAugment : Practical automated data augmentation with a reduced search space, Cebuk et al., NeurIPS 2020 )

Why need AutoML? Too LARGE search space!

\(\rightarrow\) Instead of searching, random sample

hyperparameter :

  • \(N\) : number of operations
  • \(M\) : range of operations

figure2

figure2

figure2


6) UniformAugment

( UniformAugment : A Search-free Probabilistic Data Augmentation Approach, LingChen et al. )

  • mix 2 images pixel wise ( NO SEARCH )
  • train : N-class multi-label prediction
  • Prob 0~1 of,,,
    • using certain augmentation (O,X)
    • Magnitude of augmentation

figure2


7) TrivialAugment

( Trivial Augment : Tuning-free Yet State-of-the-Art Data Augmentation, mUiller et al., ICCV 2021 )

previous methods

  • consider trade-off between efficiency & effectiveness


Proposes…

  • instead of parameter-free, just search important factors
    • (1) Augmentation Type
    • (2) Magnitude

figure2

figure2

Categories:

Updated: