Data Augmentation Strategies

Data Augmentation Strategies

CV Augmentation (albumentations)

import albumentations as A
transform = A.Compose([
    A.RandomResizedCrop(224, 224, scale=(0.2, 1.0)),
    A.HorizontalFlip(p=0.5),
    A.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.2, hue=0.1),
    A.GaussianBlur(blur_limit=(3, 7), p=0.5),
    A.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ToTensorV2(),
])

Mixing Strategies

• CutMix: replace random patch with another image's patch + mix labels
• MixUp: linear interpolation of images and labels
• Mosaic: 4 images in quadrants (YOLO-style)

Domain-Specific

• Medical: intensity windowing, elastic deformation, no flips if laterality matters
• Satellite: full rotation invariance, multi-scale crops
• Robotics: viewpoint synthesis, lighting variation, background randomization

Rules

• Never augment validation/test sets (except TTA at inference)
• Visually inspect augmented samples before training
• Match augmentation strength to dataset size (stronger for small datasets)

Key Libraries

albumentations, torchvision.transforms.v2, kornia