Automated labeling (AL) is a feature that allows you to automatically annotate the rest of your project or a specific dataset. Essentially, you take a model used in your project and apply it either to all the unannotated images in a project or a specific part of it.
Additionally, you can set the following parameters:
- Confidence threshold: the higher the confidence value, the fewer annotations you will get, but those will be the ones that our model is the most confident about;
- Images count: the number of images on which you want to run Automated labeling. Leave the field empty to apply AL to all images;
- Don't update / Update toggle: this flag determines the feature's behavior if Automated labeling did not predict a single label or tag for an image:
- If the toggle is on and the feature does not predict a single label or tag for an image, the image's status will be updated to Auto-labeled;
- If the toggle is off and the feature does not predict a single label or tag for an image, the image's status will not be updated to Auto-labeled.
Important: you can access Automated labeling (AL) once your Object Detection or Instance Segmentation assistant is trained. However, remember that the quality of the annotations produced by AL depends on the performance of the corresponding AI assistant.
Please note that there is no automated way to undo or remove labels created by Automated labeling.
- If your assistant was trained on a small number of images with few labels, probably, the labels predicted by AL will also be imprecise.
- The better the performance of your OD/IS assistant, the better will be the results of Automated labeling.
- Once the Automated labeling is performed, you can review the potential errors in labels using AI Consensus Scoring feature.