Train custom vision models for OAK, right in Luxonis Hub
Custom model training is now available in Luxonis Hub. You can prepare datasets, annotate images, train and evaluate models, and convert them for OAK—all within Luxonis Hub.
If you’re building with OAK, you may have reached a point where a general-purpose model doesn’t quite fit your application. You need to recognize your own products, distinguish specific parts, or detect objects under the lighting and camera angles of your installation. That means working with your own data.
We’re excited to make that work more accessible. Hub brings the steps together in a developer tool built to help you turn representative images into a model you can test on your camera.
From your data to a trained model
Start by uploading images or an existing annotated dataset. Hub supports classification, object detection, semantic and instance segmentation, keypoint detection, and OCR, with compatible LuxonisTrain architectures to choose from.
The workflow gives you tools to:
Prepare and label data. Annotate images in the sample editor, use Smart Annotations for supported tasks, or pre-annotate new batches with a model you’ve already trained. Review and correct the suggested labels before using them.
Keep experiments reproducible. Check dataset size, class balance, and train/validation/test splits, then create a dataset version to preserve the samples and annotations used for a run.
Train and inspect results. Choose your desired architecture, follow training progress, and review metrics and inference images to understand where the model works and where it needs better examples.
Bring the model to OAK. Successful training runs create a model version with an ONNX NN Archive and a PyTorch checkpoint. Easily convert the archive for your target platform and reference the resulting model in a DepthAI pipeline or OAK App.
Built for the next iteration
The most useful training data usually comes from the real world, after a model is already deployed: the lighting you didn't expect, the object you didn't think to include, the missed detection on a Tuesday afternoon.
With Hub, those moments don't have to get lost. Capture them as Snaps from your devices, add them to your dataset, annotate, and train the next version. The path from "the model missed this" to "the model handles this" stays inside one workflow. Your application and real camera data guide the next improvement.
Try it at no cost
Every team already working in Luxonis Hub has access to credits to try this at no cost.
Open Luxonis Hub and follow the Hub AI workflow guide to get started. Try it with data from your own application, see how the model performs, and share your results or feedback on the Luxonis forum. We’d love to hear what you build!
