What to do with wrong predictions while uploading model-based images?

I’ve just read this article https://heartbeat.fritz.ai/mobile-machine-learning-with-fritz-ai-studio-a-non-developers-journey-bb644bbede17

And I didn’t exactly understand when I upload model-based image collections what should I do with wrong predictions? I need to delete annotation ?

There are two types of annotations you can pass from the SDK to the model-based image collection. E.g.

predictor.record(image, predictedAnnotations, modifiedAnnotations)

The first are predictedAnnotations. These should be exactly what the model has predicted. They are available to view by selecting the Model tab of the annotation viewer in Fritz AI Studio.

The second are modifiedAnnotations. These appear when selecting the User tab in the annotation viewer in Fritz AI Studio.

When re-training your model ONLY User annotations are used for training. Model annotations are simple available to look at and cannot be modified.

You have a couple of options when it comes to model based image collection.

Option 1:

  1. Capture images via the predictor.record method.
  2. Pass in the exact model predictions as predictedAnnotations but leave modifiedAnnotations as nil.
  3. In the Fritz AI Studio web app, open the model-based image collection and select Annotate Images
  4. Make sure the User tab is select for an image.
  5. Annotate images as you would a Seed or Standard Image collection.
  6. Retrain your model.

Option 2:

  1. Capture images via the predictor.record method.
  2. Pass in exact model predictions as both predictedAnnotations AND modifiedAnnotations:
  3. In the Fritz AI Studio web app, open the model-based image collection and select Annotate Images
  4. Make sure the User tab is select for an image.
  5. Move any incorrect annotations to the correct location or delete any false positives.
  6. Retrain your model.

The second option may be faster if your model is producing mostly correct results.

I hope that helps!