Augmentor

Augmentor is an open-source Python library for image augmentation in machine learning, published on GitHub where it has earned more than 5,100 stars. It gives data scientists a clean, chainable API for expanding image datasets with realistic transformations, which helps computer vision models generalize instead of memorizing small training sets. The project was last actively updated in 2024 and remains a widely used reference tool for augmentation pipelines.

Overview

Instead of hand-writing transformation code with low-level image libraries, Augmentor lets you build a pipeline object, add operations with individual probabilities, and then sample as many augmented images as you need. Operations are chained together and applied stochastically, so every generated image is a fresh variation of the originals. It works with images stored on disk as well as with ground-truth masks for segmentation tasks, and it integrates naturally with Keras and other deep learning workflows. Because the library is framework-agnostic at its core, teams also use it to prepare datasets for PyTorch and TensorFlow training runs, and its small API surface makes it a common first step in computer vision courses and production data pipelines alike.

Key Features

• Pipeline-based API: add operations, set their probabilities, and sample augmented images in a few lines of code
• Wide operation set including rotation, skewing, shearing, cropping, zooming, elastic distortion and perspective transforms
• Color and contrast adjustments such as brightness, hue and histogram equalization
• Ground-truth image support so segmentation masks are transformed identically to their source images
• Multi-threaded sampling for generating large augmented datasets quickly
• Works with standard image formats and directory-based dataset layouts

💳 Pricing

Free and open source under the MIT license. There is no paid tier, hosted version or commercial edition.

Who Is It For

Machine learning engineers and researchers training image classifiers, detectors or segmentation models who need larger, more varied training sets without collecting more photos. It is especially handy for small datasets where overfitting is the main risk.