Autotag is a powerful machine learning feature for adding refined visual similarity scores as metadata to images and/or objects within your dataset. You start with an initial set of example items to seed your search, then iterate through multiple refinement steps where you finetune the search by identifying relevant items from the pool of returned search results. Once you're finished finetuning your search, you can commit the Autotag - this adds search scores (the higher, the more relevant the item) as metadata to the top twenty thousand most relevant items in your dataset. You can then query on these items within the Nucleus grid dashboard, or export them via API.
An example usecase for Autotag is if you found that your model performed worse on trucks at night, and wanted to label more images with trucks at night to further train your model. With a dataset of unlabeled images, you can use Autotag to search for all your images that contain trucks at night. You can then select the resulting items and directly send them to Scale labeling!
* This applies to all datasets created after November 30th, 2021. If your dataset was created before this date and you would like to enable Autotag and Similarity Search with internal embeddings calculated by Nucleus, you need to turn on continuous indexing for your dataset via our API.
Updated 11 months ago