imgaug

The library supports geometric transforms (affine transformations, perspective warp, elastic deformation, piecewise affine), color-space operations (brightness, contrast, gamma, hue, saturation), convolutional effects (blur, sharpen, emboss, edge detection), and noise injection (Gaussian noise, salt-and-pepper, dropout).

Crucially, imgaug correctly propagates augmentations to associated annotations bounding boxes, segmentation masks, heatmaps, and keypoints all transform consistently with the underlying image.

A distinctive feature of imgaug is its powerful augmentation pipeline composition system.

Transforms can be combined with Sometimes (probabilistic application), OneOf (random selection from a group), SomeOf (applying N of M transforms), and Sequential (deterministic ordering), enabling complex, realistic augmentation policies that closely mimic natural image variability.

The library supports both deterministic and stochastic modes: deterministic mode allows applying the exact same augmentation to multiple related images (useful for video frames or multi-modal data), while stochastic mode generates varied outputs for standard supervised training.

imgaug has been widely adopted in academic research and competition settings for tasks including object detection (PASCAL VOC, COCO-style), medical image segmentation, remote sensing analysis, and document understanding.

Its thorough documentation and extensive example gallery make it accessible to practitioners new to data augmentation, while its fine-grained control over augmentation parameters satisfies the needs of domain experts who need to model specific real-world degradations.