Moving from Azure Machine Learning data labeling

Microsoft retired Azure Machine Learning data labeling on 30 September 2026 and asks its users to move to a third-party labeling tool. Its notice says that workloads still running data labeling are deleted and their application data lost (Microsoft Learn).

Annotide reads the images where Azure Machine Learning left them, in your Blob Storage, so nothing is copied. It imports the COCO files that data labeling exported, and it signs people in with Microsoft Entra ID.

What carries over

Azure ML project type	Export	In Annotide
Object identification (bounding box)	COCO	Imported: boxes, with your class names
Instance segmentation (polygon)	COCO	Imported: polygons
Image classification (multi-class, multi-label)	CSV, MLTable	Not imported yet: re-label, or pre-label with your own model
Semantic segmentation (preview)	MLTable	Not imported yet
Text classification, named entity recognition	CSV, MLTable, CoNLL	Not imported yet

The images themselves never move: Annotide’s Azure Blob connector points at the same storage account and container the Azure ML datastore used.

What you need

• The COCO export of each project. Azure ML wrote it to the workspace’s default Blob store, in a folder under Labeling/export/coco. Projects that were never exported before the retirement may be gone.
• The storage account holding the images. That is often the workspace’s own (workspaceblobstore).
• An Annotide installation: the Azure Marketplace offer, or the Helm chart.

Steps

1. Connect the storage. On Connectors, register an Azure Blob connector for the account and container that hold the images. Use the managed identity, not a key, and run Check: browsers load images straight from the storage, so it needs CORS for the app’s origin. A storage account that only has a private endpoint (common for Azure ML workspaces) works only for people on that network (INSTALL.md §5).
2. Create the project. Set its folder (prefix) to where the images are, the same path the Azure ML dataset used.
3. Recreate the classes in the label schema with the bounding box or polygon tool. Use the Azure ML label names, or map them during the import.
4. Scan the source (project page → Add data). Scanning reads every image’s width and height. The import needs them, because Azure ML’s COCO export stores coordinates as fractions of the image and writes 0 for every width and height.
5. Import. In Imports, choose COCO, pick the exported file (upload it, or name its path on a connector) and tick Dry run first. The result shows how many images matched project items, which ones did not, and the class names found, so you can map them before the real import.
6. Import for real as submitted (ready for review) or draft. Each image gets a new annotation version, authored by you.

Annotide reads Azure ML’s image references (azureml://…/datastores/<name>/paths/<path>, AmlDatastore://<name>/<path> or a Blob URL) as the <path> inside them. When the project’s folder differs, images still match by their file name as long as that is unique in the project.

Replacing the rest of Azure ML data labeling

Azure ML data labeling	Annotide
ML-assisted labeling (clustering, prelabeling)	Pre-labelling with your own models: an endpoint that speaks Annotide’s model API, or pre-labels posted through the SDK (from an Azure ML job, for example); the queue can be ordered by uncertainty
Consensus labeling (preview)	Consensus: several annotators per item, with an agreement score; plus gold items
Review labels, accept and reject	Review queue: approve, reject with a comment, or correct
Export to COCO, CSV or MLTable	Exports to COCO, YOLO, spaCy, CoNLL and the native format; snapshots with train / val / test splits; MLflow for Azure ML training
Workspace roles	Entra ID single sign-on with group-to-role mapping (ENTRA.md)
Vendor labeling companies	Any team you invite, limited to folders if needed

Help

Migrating a large project, or a format not imported yet? Write to hello@annotide.com with the project type and roughly how many items it has.