( 참고 : 패스트 캠퍼스 , 한번에 끝내는 컴퓨터비전 초격차 패키지 )
Data Augmentation (1)
Categories of Data Augmentation
- (1) Rule-based
- (2) GAN-based
- (3) AutoML-based
(1) Rule-based
(1-1) Color Transformation
- Gaussian Blur, Motion Blur
- Brightness Jitter, Contrast Jitter, Saturation Jitter
- ISO Noise
- JPEG compression
(1-2) Spatial Transformation
- Flip, Rotation, Crop, Affine …
(1-3) ETC
- a) Data Mixing / Erasing
- b) PatchSuffle Regulaization
- c) Sample Pairing
- d) Mixup
- e) Mosaic Augmentation ( Cropping and Patching )
- f) Multiple Way of Mixing
- g) Manifold Mixup
- h) Random Erasing / Cutout
- i) Hide-And-Seek for Weakly-Supervised Localization
- j) CutMix
- k) AugMix
- l) SmoothMix
a) Data Mixing / Erasing
- train by N-label prediction task
b) PatchSuffle Regulaization
- shuffle pixels inside window
- cons) need hyperparameter \(N\) ( = window size )
c) Sample Pairing
- mix 2 images pixel-wise
- training : by \(N\)-class multi-label prediction
d) Mixup
- mix 2 images, with linear interpolation
e) Mosaic Augmentation ( Cropping and Patching )
-
gather multiple patches from multiple images
-
solve N-label classification task
f) Multiple Way of Mixing
- mix 2 images pixel-wise
- propose 8-ways of mixing
g) Manifold Mixup
- Mix up in the hidden representation space!
- more natural/smooth boundary
h) Random Erasing / Cutout
-
erase randomly ( black / white / random … )
-
not only “random erasing + random cropping”,
but also “image/object-aware erasing”
-
used in object detection & person re-identification
i) Hide-And-Seek for Weakly-Supervised Localization
-
gather multiple patches from multiple images
-
task : object localization
( = finding bounding box )
j) CutMix
- cropping + patching
- task : classification, weakly-supervised object localization
k) AugMix
- mix augmentations!
- with different weight per augmentation methods!
l) SmoothMix
- similar to CutMix, but when patching, smooth the boundary!
