Use the segmentation model catalog

Choose the segmentation outcome first#

Choose the task from the result you need, not from a model name. State that outcome clearly in the training instruction so planning can select compatible targets, architecture, loss, metrics, and transforms.

• Semantic segmentation assigns a class to each region or pixel without preserving separate objects of the same class.
• Instance segmentation preserves separate masks for individual objects, including touching objects of the same class.
• Panoptic segmentation combines instance-aware objects with background or other non-instance regions in one structured result.

Understand the new catalog families#

• EoMT DINOv2 Small 640, Base 640, and Large 640 are catalog choices for panoptic segmentation.
• RF-DETR Segmentation Nano, Small, Medium, and Large are catalog choices for instance segmentation.
• RTMDet-Ins Tiny, Small, Medium, and Large are catalog choices for instance segmentation.

Review or change the planned architecture#

1. Let planning choose the initial architecture

   Create the pipeline from the dataset and natural-language instruction. The platform does not require a model choice in the creation dialog.

2. Wait for planning to complete

   Do not start the run until the generated task and resources are available for review.

3. Open the Architecture resource

   In Pipeline Tasks, expand the model-training task and open its Architecture resource.

4. Open Knowledge Base

   Review Related Architectures for the current task. A family that is absent is not an available replacement for this planned resource.

5. Check task fit and availability

   Confirm the candidate’s segmentation target, variant, parameters, availability indication, and support on the selected ML backend before choosing Use This Architecture.

6. Save before the first run

   Save the replacement and reread downstream resources before running. After a pipeline has run, clone it first when a different architecture is required.

Verify the contract before and after the run#

• Before running, confirm the selected dataset, segmentation target, architecture, backend, loss, metrics, transforms, input handling, and validation warnings.
• For panoptic work, confirm that labels distinguish individual objects from non-instance regions where the dataset requires both.
• For instance work, confirm that annotations preserve separate objects rather than only a merged class mask.
• Interpret only the metrics and result fields returned by the completed model version; do not compare semantic, instance, and panoptic measures as if they were interchangeable.
• Record the exact pipeline and model version when reviewing, reporting, or deploying a result.