Tips for Training your own AI Models
This guide covers best practices for getting the most out of your ONE AI training runs. For basic setup and your first project, see the Quick Start Guide. For dataset handling details, see the Dataset Guide.
Add filters to process your images
🔗 guide on image filters
You can use filters to prepare your images for the training process. To get the best results, you should try to simplify the images as much as possible, while keeping all relevant information for your application. For example, you should use the Resolution Filter to reduce the image size to the minimum that still keeps the important details visible. This makes the task a lot easier for the AI model, because it needs to analyze a smaller image area to make its predictions. Other useful filters for reducing unnecessary information are the Crop Filter, which crops the image to the region of interest or the Color Filter combined with the Channel Filter to convert the images to grayscale and only keep a single color channel.
You can also use filters to enhance relevant information in your images. For example, you can use the Color Filter to adapt the contrast or brightness of your images, making it easier for the model to distinguish between different features.
Add augmentations to increase dataset size
🔗 guide on augmentations
You can use data augmentation to artificially increase the size of your dataset by applying random transformations to your images during the training process. This can help to improve the performance of your AI model and prevent overfitting, especially if you have a small dataset. You can also use augmentations more targeted to simulate specific conditions that might occur in your application, such as different lighting conditions or rotations.
The default augmentations are a good starting point, but we advise to tune them to your specific application. You should keep in mind that not all augmentations are suitable for every use case. For example, if you are trying to detect labels, it doesn't make sense to flip the images, because your model won't encounter a mirrored label in the real world.
Provide context for your application
To get the best results, you should provide some context information about your application in the Model Settings tab. By providing some simple estimates, like the general task complexity or the variance between objects of the same class, ONE AI can tune the model architecture to your specific use case.
Evaluating your Model
You can run the model test to compute various performance metrics for your trained model. This will also visualize the predictions of your model for a couple of test images. For further evaluation, you can export your model as an ONNX file and run it directly in OneWare Studio. You can add it in the Annotation Tool, to make predictions for individual images or use it in the Camera Tool to test it with a live camera feed.
Where to go from here?
We have several different guides and tutorials to help you get started with ONE AI:
Need Help? We're Here for You!
Christopher from our development team is ready to help with any questions about ONE AI usage, troubleshooting, or optimization. Don't hesitate to reach out!