Create new project reusing existing dataset #4988
Description
Users often want to experiment with different tasks and labels, while also reusing previously created datasets. To facilitate this process, we want to enable the creation of new projects from existing datasets.
In Geti Classic, a similar usecase was addressed through dataset import/export: users had to export their datasets to Datumaro format, and later reimport them to create a new project. In Geti Tune, we want to avoid the intermediate I/E steps, all data should stay inside the application during this operation.
Requirements
- The user should have an option to create a new project initialized with the dataset of another project
- The dataset should be a copy, so that operations on one project don't affect the other (cross-project isolation)
- The new project may have a different task type - see below for the list of combinations to support. Annotations must be converted to the target type, as described below.
- The new project may have different labels than the original one. The user should have the option to map each source label to a target label, or none (in this case, annotations with that label are discarded). It is allowed to assign two source labels to the same target label.
Supported task conversions
| From | To | How to convert shapes |
|---|---|---|
| Instance segmentation | Object detection | Create the smallest bounding box enclosing the polygon |
| Instance segmentation | Multilabel classification | Apply labels of all polygons in the image |
| Object detection | Instance segmentation | Reinterpret bbox as polygon |
| Object detection | Multilabel classification | Apply labels of all polygons in the image |