Skip to main content
Augmentation squeezes more value out of the images you already have — cleaning up your training split and generating realistic variations so your model generalizes better. You configure it while creating a dataset version.
Preprocessing and augmentation apply to the training split only. Your validation and test sets stay untouched, so evaluation stays honest.

​
Preprocess

Preprocessing prepares images before training. The most common step is Filter Null — dropping images with no annotations (or capping how many empty images to keep), so the model isn’t diluted by blank frames.

​
Augment

Augmentation applies stochastic transforms to the training split, multiplying it into new, realistic examples. It’s powered by Albumentations, so its full library of image transforms — flips, rotation, blur, noise, brightness, and more — is available. Add one or more steps:
  1. In the snapshot’s Augmentation section, select Add Augmentation Step.
  2. Choose a transform and set its parameters (for example, a 50% chance of a horizontal flip).
Augmentation steps

Stack transforms to mirror the variation your model will see in production.

Match augmentation to reality: if parts arrive at any rotation, add rotation; if lighting varies, add brightness. Don’t add transforms your camera will never see.

​
Next steps

Train a model

Put your dataset version to work and train a detector.

Create a dataset version

Where preprocessing and augmentation are configured.