From photos to a tested model in four steps

1Build a dataset

Create a dataset and name the things the model should learn to find, such as 'bolt' or 'scratch'. These are your classes. Then mark them in your images: draw a box around each object, trace its exact outline (polygon), or place points on fixed spots of it (keypoints). For outlines, SAM 2 helps: click the object once and the app draws the outline for you to adjust.

You get a labeled set of images, split automatically into images to train on and images to test with.

• Auto-label. Once you have a trained model, let it label new images for you. Its suggestions arrive as drafts in the labeling screen; you check them, fix what is wrong and save.
• Augmentation. Add flipped, rotated or brightened copies of your images, so the model learns from more examples.
• Auto-balance. If one class has far fewer examples than the others, the app adds extra cut-outs of it to the training images.
• Label Studio. Already label in Label Studio, a separate labeling program? You can import its labels instead.

2Train a model

Pick your dataset and a model size, and press Start Training. Training runs on your own GPU. You can stop it and continue later from where it stopped.

You get a trained model, with a score and a chart that show how well it is learning. A built-in tour explains what the score means.

• Three kinds of model. One finds objects and draws a box around them (Faster R-CNN). One also places keypoints on them (Keypoint R-CNN). And one traces the exact outline of each object (Mask R-CNN).
• Three sizes. A larger model can be more accurate but needs more VRAM. The app shows how much before you start.

3Test the model

Run the model on test images it did not see during training, or on photos of your own.

You get the images with what the model found, a score per class, and a confusion matrix: a table that shows which classes the model mixes up, with its biggest mistakes listed underneath. Outline models are scored on their outlines, not just on a box.

4Compare models

All your tested models stand in one ranking, best score on top. Pick two of them, for example one from before and one from after you added more images.

You get their scores and results side by side, so you can see which one is better and keep that one.