Augmentation
Create training-only variations without changing original images or annotations.
Before you start
Choose a dataset and expand Augmentation. It is Off by default.
Quick start
- Enable only variations that preserve your classes. Brightness and contrast are selected initially; rotation and mirroring are opt-in.
- Choose a Train image and inspect Original beside Augmented. New example changes the sample variation.
- Check boxes or masks in the preview. Disable unsuitable transformations before starting training.
Functions and choices
Choose variations
Rotation uses 90° increments. Mirroring may change symbols or labels; keep it Off when orientation matters, including suitable dice tasks.
Preview
The image picker only selects a preview. All Train images participate; Val, Test and Guard images are not augmented by this policy.
Result
A saved augmentation policy for the run.
Next step
Start training, then judge the result on unchanged validation images.