Hierarchical classification is no longer supported in Geti v3.0

[yvolokitin]

We want to inform the community about a change to supported workflows in the current release.

Hierarchical classification (h-cls) projects are notoriously challenging to work with — not only because the label structure is hard to set up, but also due to the general unavailability of annotated datasets with hierarchical relationships. Moreover, such relationships have to be enforced when post-processing the labels predicted by a model. Practically speaking, this complexity is hard to justify, so we decided to discontinue the hierarchical classification task in Geti v3.

What to do instead
From Geti v3, users should model their h-cls problems as a simple multiclass classification, defining one label for each leaf of the original hierarchy.

Example — a hierarchical classification problem with this class hierarchy:

can be remodeled as a multiclass classification problem with only the leaf categories:

Removing the parent labels isn't expected to cause side effects at training time, nor a performance drop in the resulting model compared to v2.

Migrating an h-cls project to v3

If you have a hierarchical classification project in Geti v2 and want to migrate it to Geti v3 (as a multiclass project), follow these steps:

1. Export the dataset from your Geti v2 instance in VOC format.
2. In Geti v3, choose "Create new project from dataset".
3. Upload the zip archive.
4. Select "Multiclass classification" as task type.
5. In the label selector, select only the leaf labels.
6. Click Create and wait for the import to finish.
7. Verify that all media and their annotations are correctly imported.

Note: The export format must be VOC — the Datumaro format does not support this scenario.

Feedback
If hierarchical classification is important to your use case, please share your label structure and requirements in the comments so we can factor it into future planning.

📖 For the full step-by-step guide, see the v2 → v3 migration documentation