Source-linked AI summary
Albumentations: fast and flexible image augmentations
Alexander Buslaev, Alex Parinov, Eugene Khvedchenya, Vladimir I. Iglovikov, Alexandr A. Kalinin
TL;DR
Limited labeled data and restricted transformation choices motivate a flexible image-augmentation library. The paper presents Albumentations, which supports varied task-specific transformations and wraps other libraries; it reports faster processing than commonly used tools for most common transformations.
Problem
Limited labeled datasets increase overfitting risk, while existing frameworks typically offer only basic image transformations and vary in augmentation speed.
Method
Albumentations provides many image transformations through a simple interface for classification, segmentation, and detection, while also wrapping other augmentation libraries.
Results
Albumentations is faster than other commonly used image augmentation tools on most commonly used image transformations.
Takeaways & Limitations
Albumentations offers a fast, flexible augmentation library with task-specific examples and publicly available source code.
Abstract
from arXiv · showhide
Data augmentation is a commonly used technique for increasing both the size and the diversity of labeled training sets by leveraging input transformations that preserve output labels. In computer vision domain, image augmentations have become a common implicit regularization technique to combat overfitting in deep convolutional neural networks and are ubiquitously used to improve performance. While most deep learning frameworks implement basic image transformations, the list is typically limited to some variations and combinations of flipping, rotating, scaling, and cropping. Moreover, the image processing speed varies in existing tools for image augmentation. We present Albumentations, a fast and flexible library for image augmentations with many various image transform operations available, that is also an easy-to-use wrapper around other augmentation libraries. We provide examples of image augmentations for different computer vision tasks and show that Albumentations is faster than other commonly used image augmentation tools on the most of commonly used image transformations. The source code for Albumentations is made publicly available online at https://github.com/albu/albumentations
I. INTRODUCTION
Deep vision models risk overfitting when labeled data are limited, motivating augmentation that expands training-set size and diversity. Existing tools often provide limited transformation choices and variable processing speed, so the paper presents Albumentations as a faster, more flexible alternative.
- I. INTRODUCTION: Large neural networks can overfit when massive labeled datasets are unavailable.Their many parameters support generalization with abundant labeled data, but limited data elevate overfitting risk.
- I. INTRODUCTION: Data augmentation increases labeled training-set size and diversity through input transformations that preserve output labels.
- I. INTRODUCTION: Most deep learning frameworks limit image transformations largely to flipping, rotating, scaling, and cropping.Different domains, imaging modalities, and tasks may require broader combinations.
- I. INTRODUCTION: Rotation improves CIFAR-10 augmentation but can harm MNIST digit discrimination between 6 and 9.
- I. INTRODUCTION: Albumentations addresses the importance of augmentation speed because preprocessing typically runs on CPUs while neural-network computation increasingly uses GPUs.The paper reports that processing speed varies among existing augmentation tools.
- I. INTRODUCTION: Albumentations combines a broad range of transformations with a simple interface for classification, segmentation, and detection.It also wraps other augmentation libraries.
II. ALBUMENTATIONS
Albumentations offers many image transformations through an easy-to-use interface, illustrated with task-specific examples and visually distinct combinations of operations.
- II. ALBUMENTATIONS: Albumentations provides color, contrast, brightness, and other image transformations, including combinations of these operations.The example parameters were chosen so results visibly differ from the original image, which may not always be desirable.
- II. ALBUMENTATIONS: The library presents task-specific augmentation examples for different computer vision tasks.
- II. ALBUMENTATIONS: Figure 1 illustrates exemplar applications of image transformations available in Albumentations.
A. Street view image detection
For street-view vision, Albumentations applies spatial transformations jointly to images and structured annotations used in detection, tracking, and segmentation.
- A. Street view image detection: Many spatial transforms operate on segmentation masks and bounding boxes as well as images.These annotation formats support object detection and tracking, while masks support segmentation.
- A. Street view image detection: A Mapillary Vistas example combines a horizontal flip and random-sized crop across an image, bounding box, and instance-mask annotations.
B. Satellite and aerial imagery
Satellite and aerial imagery commonly uses shape-preserving transformations to avoid distorting rigid objects such as buildings. Albumentations illustrates these operations in aerial-image applications.
- B. Satellite and aerial imagery: Satellite and aerial-image workflows use cropping, rotations, reflections, and scaling to preserve object shapes.The motivation is avoiding distortions of rigid-shape objects such as buildings.
- B. Satellite and aerial imagery: A top-three DSTL Satellite Imagery challenge solution combined random cropping with a random transformation from dihedral group Dih4.
- B. Satellite and aerial imagery: Figure 2 depicts transformations applied to an original image together with bounding boxes and ground-truth masks for instance segmentation.
C. Biomedical image analysis
Medical image augmentation is widely used because labeled datasets are typically limited. Prior work reports benefits for detection and segmentation tasks, including settings with very small image collections.
- Medical image augmentations are intensively used because labeled data are typically limited.
- Multiple augmentation techniques combined with pretrained networks enabled accurate breast-cancer detection from fewer than 100 histology images per class.
- Medical image augmentations helped improve hand-radiograph segmentation and bone-age assessment.
III. BENCHMARKS
The benchmarks compare augmentation-tool processing speed and illustrate transformations used across imaging tasks. Albumentations is reported as consistently faster than the alternatives for most image operations.
- Albumentations is consistently faster than imgaug, Keras, and PyTorch for most image transformation operations.The comparison is based on time per image transformation operation, with lower values indicating better speed.
- The benchmark section includes geometry-preserving transforms for satellite images and masks, plus grid distortion and elastic transforms for medical images.
- Table I reports time in seconds per image transformation operation task, where lower values are better.
IV. CONCLUSIONS
The paper concludes by presenting Albumentations as a fast, flexible augmentation library with many transforms and an easy-to-use wrapper around other libraries. It provides task examples and is reported faster than commonly used alternatives for most common transformations.
- Albumentations combines many image transform operations with an easy-to-use wrapper around other augmentation libraries.
- The paper provides augmentation examples for different computer vision tasks and reports faster processing than commonly used tools for most common transformations.