albumentations

It provides a comprehensive collection of over 70 augmentation transforms geometric transformations (rotation, perspective distortion, elastic deformation), color-space manipulations (brightness, contrast, hue, saturation), blur and noise operations, and domain-specific transforms for medical imaging, satellite imagery, and scene understanding.

Albumentations is built on top of NumPy, OpenCV, and imgaug, with careful optimization ensuring augmentation pipelines run faster than comparable libraries, a critical advantage when augmentation is a bottleneck in GPU training loops.

The library's API is designed around composable pipelines where transforms are chained and applied probabilistically.

A typical pipeline might randomly apply horizontal flips, random crops, one of several blur operations, and brightness/contrast jitter with each transform having an independent probability of activation per image.

This stochastic composition drastically expands the effective training set size, improving model generalization on tasks like object detection, semantic segmentation, instance segmentation, and keypoint estimation.

Albumentations correctly propagates augmentations to associated labels: rotating an image also rotates its bounding boxes, segmentation masks, and keypoints.

Albumentations integrates with all major deep learning frameworks (PyTorch, TensorFlow/Keras, JAX) and is compatible with popular computer vision libraries (MMDetection, Detectron2, YOLO variants).