Labelling the images
====================

An object detector learns from examples that are labelled: each
training image needs a box around every target object, tagged with
its class. Labelling hundreds of frames by hand is slow, so Roboflow
automates it.

Auto Label
----------

On the **Annotate** page, *Auto Label* drives a text-prompted
foundation model: you describe each class in words and it finds and
boxes those objects across the whole batch. Add a class per thing
you want to detect -- ``stuffed raccoon toy``, and ``person`` to
teach the model what to ignore -- preview the result on a few test
images, and adjust each class's confidence threshold until the boxes
land where they should.

.. figure:: figures/auto-label.jpg
   :class: framed
   :width: 100%
   :alt: Roboflow's Auto Label page: text-prompted classes with confidence sliders on the left, and a preview image with a person and a stuffed raccoon toy detected and masked

   Auto Label finds the classes from text prompts and
   labels the batch -- preview and tune the thresholds
   before running it on every image.

Run it on the batch, then review: scan the labelled images, fix the
few the model got wrong, and delete boxes it invented. Auto Label
does the bulk work; the review pass catches its mistakes.

Adding to the dataset
---------------------

Labelled images move into the dataset with a *train / valid / test*
split. The split is how the model's accuracy is measured: it trains
on the training images, tunes against the validation set, and is
scored on the test images it never saw during training. The default
split works -- accept it and the dataset is ready to train.