Skip to content

Workflow

Datasets & augmentation

Once your images are annotated, turn them into a dataset your model can learn from. VisoLabel lets you review everything in a gallery and generate augmented copies to make a small dataset go much further — all locally.

The dataset gallery

Open a project to see every image in the gallery. Annotated items are marked so you can tell at a glance what's ready and what still needs work. The left panel lists your datasets — the original Current Project plus any augmented datasets you generate.

VisoLabel dataset gallery showing a grid of annotated images, a datasets panel on the left, and Create Augmented Dataset and Train buttons.
The project gallery. Every image shows its annotation status; the left panel switches between the original and augmented datasets.

From here you can jump straight to Train, or click Create Augmented Dataset to expand your data first.

Why augment?

Augmentation creates new training examples by transforming the images you already labeled — rotating, flipping, changing colors, and so on. Your annotations are transformed along with the images, so you get more variety for your model to learn from without labeling anything new. It's the easiest way to improve accuracy when you have a limited number of images.

Creating an augmented dataset

The Data Augmentation dialog walks you through it. Give the dataset a name, choose how many augmented copies to make per image, and pick which transformations to apply. VisoLabel shows you the resulting size up front — for example, 2,000 original images at a ×3 multiplier produces 6,000 training images.

VisoLabel Data Augmentation dialog. Left side shows dataset name, an augmentations-per-image summary (2000 originals times 3 equals 6000 output), and output location. Right side shows preprocessing cards: Tile, Resize, Grayscale, Remove Empty Images.
The augmentation dialog. The multiplier controls how many variants each image produces; the live count shows your total output dataset size.

Augmented datasets are written to your local VisoLabel data folder (…/VisoLabel/datasets/augmented) — nothing leaves your machine.

Preprocessing

Preprocessing runs before the random augmentations and applies to every image. Click a card to toggle it on or off:

  • Tile — split large images into smaller tiles.
  • Resize — standardize images to a fixed size.
  • Grayscale — drop color when it isn't relevant to your task.
  • Remove empty images — exclude images that have no annotations.

Image-level augmentations

These are the random transformations applied to create each new variant. Enable as many as make sense for your data:

VisoLabel image-level augmentation cards: Rotation (enabled), Flip (enabled), Shear, Grayscale, Hue, Saturation, with a Create Dataset button.
Image-level augmentations such as Rotation, Flip, Shear, Hue, and Saturation. Enabled cards are highlighted.
  • Rotation — rotate the image by a random angle.
  • Flip — mirror horizontally or vertically.
  • Shear — skew the image to simulate perspective.
  • Hue & Saturation — vary colors to handle different lighting.
  • Grayscale — randomly desaturate some copies.

Tip: more augmentation isn't always better. Only enable transformations that reflect real-world variation in your data — for example, skip vertical flips if your objects are never upside down.

When you're happy with the settings, click Create Dataset. The new augmented dataset appears in the gallery's dataset panel, ready to select when you train a model.

Next: Training — train on your dataset locally or on Colab →