Wrap-up and next steps

Your webcam reads dice with a model you trained on your own images, and a module running on your machine turns what it reads into Yahtzee scores. You also know, from measurements rather than a guess, which of two models does the job better.

What you built

• Four datasets: a hand-labeled starter set, an auto-labeled training set, and clean and hard test datasets.
• A starter model that labeled most of your training images for you.
• Two dice detectors, one from built-in training and one from a YOLO training script, scored the same way on the same test datasets.
• A deployed vision service running the winning model on your machine.
• An inline module that reads a settled roll and scores it.

What you learned

• Add images from a camera to a dataset.
• Label images, and use a starter model to label more of them for you.
• Train an object detection model with built-in training.
• Train a YOLO model with a training script from the registry.
• Evaluate models on held-out test datasets and compare the results.
• Choose a confidence threshold from evaluation results.
• Deploy a model with the ML model and vision services.
• Build an inline module that scores rolls from your model’s detections.

What’s next

• Automate annotation: auto-label new images with your best model, add images from the conditions where it struggled, retrain, and evaluate the new version on the same test datasets.
• Custom training scripts: write your own training script, or fork yolo-training and try a larger imgsz to help with small pips.
• Configure a vision service: see the other kinds of vision services, such as classifiers and segmenters.
• Miniature Palletizing with the SO-ARM101: give your machine an arm, and you’re one step from a robot that rolls its own dice.

To keep going with the scorer, add a command to your module that keeps a scorecard across 13 rounds, so you can play a full game against your machine.