Data Augmentation
The short answer · 14 words
In short
Data augmentation generates task-consistent variants of existing training examples to train models more robustly.
Flow-alerta.cab (2026): Data Augmentation. https://flow-alerta.cab/en/glossar/data-augmentation. Editorially reviewed on: 10 October 2026.
How much detail would you like?
Practical
Augmentation can reduce over-adaptation and improve robustness against expected variations. It does not replace representative data and does not guarantee better generalisation.
Product images are mirrored, cropped and changed in color within realistic limits; bounding boxes are also adapted.
Companies derive permitted transformations from real use and check effects on subgroups. SMEs start with few plausible variants and a clean baseline comparison.
At a glance
- Core
- In data augmentation, training data is extended by transformations that generate meaningful variants for the task. In the case of classification, the label often remains the same; in the case of object recognition or segmentation, annotations must be appropriately transformed.
- Use
- Augmentation can reduce over-adaptation and improve robustness against expected variations. It does not replace representative data and does not guarantee better generalisation.
- Example
- Product images are mirrored, cropped and changed in color within realistic limits; bounding boxes are also adapted.
- Limits
- Unrealistic transformations, false label assumptions, application to test data or before the train test split. Data enrichment, oversampling and fully synthetic data are other concepts.
In context
Questions about this term
Should labels remain the same?
Not always; structured annotations often have to be transformed.
Can the test data set be augmented?
For the regular evaluation, generally not as a training substitute.
Is synthetic data generation the same?
No.
All details and background
In data augmentation, training data is extended by transformations that generate meaningful variants for the task. In the case of classification, the label often remains the same; in the case of object recognition or segmentation, annotations must be appropriately transformed.
ELI5 – Explained simply
From a learning image, rotated, tailored or slightly modified variants are created, so that the model does not only know one view.
Why it matters in practice
Augmentation can reduce over-adaptation and improve robustness against expected variations. It does not replace representative data and does not guarantee better generalisation.
Example
Product images are mirrored, cropped and changed in color within realistic limits; bounding boxes are also adapted.
Pitfalls and limits
- Unrealistic transformations, false label assumptions, application to test data or before the train test split. Data enrichment, oversampling and fully synthetic data are other concepts.
Origin & history
Augmentation has been disseminated especially with computer vision and deep learning, but also exists for text, audio and time series. Modern frameworks integrate transformations directly into training pipelines.
Use in mid-sized companies
- Companies derive permitted transformations from real use and check effects on subgroups. SMEs start with few plausible variants and a clean baseline comparison.
Related terms
- Overfitting
- Synthetic data
- Computer vision
- Train test split
- Robustness
Matching Flow-alerta service
AI opportunity analysis & roadmapAI data preparation: Develop appropriate transformation rules, leakage-safe pipeline and robust efficacy comparison.Matching tool / use case
View use casesTensorFlow/Keras or Albumentations; Use Case: Train image classification with realistic perspective and light variants.Sources
- www.tensorflow.orgTensorFlow, „Data augmentation“ (Abruf 05.10.2026)
Quellenprüfung: 10.10.2026. Definition und dokumentierte Einordnung abgeglichen; Beispiele beschreiben mögliche Anwendungen.
Editorially reviewed on: