Augmentation

Augmentation is the practice of creating modified copies of training images to artificially expand the dataset and make models more robust. Instead of collecting and labeling thousands more images, you apply random transformations to existing ones: horizontal flips, rotations, crops, brightness changes, blur, noise injection, and color shifts. The model sees each image in many different forms, which reduces overfitting and improves generalization to real-world conditions.

Geometric augmentations (flip, rotate, scale, translate, shear) teach the model that objects can appear in different positions and orientations. Photometric augmentations (brightness, contrast, saturation, hue, gaussian noise) teach robustness to lighting and sensor variations. Advanced techniques like Mosaic (stitching four images into one), MixUp (blending two images and their labels), and CutMix (pasting a patch from one image onto another) create more complex training signals. Copy-Paste augmentation places object instances from one image into another, which is especially effective for instance segmentation.

Most training frameworks apply augmentations on-the-fly during data loading, so they don't increase disk usage. The key is matching augmentations to your deployment scenario: if your camera is always upright, vertical flips add noise rather than value. If lighting varies heavily, aggressive color augmentations help more than geometric ones.