Phase 2: Label the starter set

At the end of this phase, every die in dice-starter has a box around it labeled with its value, at least 10 times for each face. That’s the minimum a training job accepts, and it’s what the starter model in Phase 3 learns from. This is the only time you label a whole dataset by hand, so it’s worth being careful: about 100 boxes, one die at a time.

Draw a box on every die

Use the labels 1, 2, 3, 4, 5 and 6, for the number of pips on the top face.

  1. On the DATASETS page, open dice-starter and click the first image.
  2. In the side panel, click the Actions tab, then click Annotate.
  3. Select or create the label for the first die’s top face, for example 4.
  4. Hold Cmd (macOS) or Ctrl (Linux) and drag a rectangle tightly around the die.
  5. Repeat for every die in the image, then move to the next image.

For more on the annotation tools, see Annotate images.

Follow these rules on every image, here and in later phases:

  • Box the whole die, edge to edge, with as little background as possible.
  • Label the face pointing at the camera, not a side face.

Your boxes should look like these, with each one tight around a die, including dice that touch:

Three images from a labeled starter dataset, with a tight box around every die and each box labeled with the die's top face: four dice spread apart, three dice in a diagonal line, and four dice touching in a cluster.

Checkpoint

Every image in dice-starter shows a labeled box on each die. If you find a die with no box, go back and box it; otherwise the model learns that a die can be background.

Check the label counts

  1. On the DATASETS page, open dice-starter.
  2. In the dataset’s side panel, find Bounding box labels. Each row shows how many images use the label and, in parentheses, how many boxes have it.
The dice-starter dataset page with 30 images. The side panel's Bounding box labels list shows labels 1 to 6, each with at least 10 boxes, from 3 with 17 images and 23 labels down to 4 with 10 images and 13 labels.

The Unlabeled count under Classifications and tags counts images with no classification tag. This workshop uses bounding boxes, not tags, so that count stays at 30. It doesn’t affect training: when you train an object detection model, the 20% unlabeled limit counts only images with no box for any label you train on, and classification tags don’t count.

Checkpoint

Each label, 1 through 6, has at least 10 boxes, and Bounding box labels shows no Unlabeled row. If you see a label you didn’t mean to create, such as l or six, relabel its boxes and delete it. If a face has fewer, go back to the Test section of the camera card on the CONFIGURE tab, add a few images where that face is on top, and label them. A training job won’t start with fewer than 10 boxes for any label you train on.

In the next phase, you train a starter model on these 30 images, and capture the rest of your training images while it trains.