Start with the classes your task needs
A label scheme should make the target task easier to define and evaluate. A navigation task may need obstacle classes; a manipulation task may need the drawer front and handle to be distinguishable. More labels are useful only when the pipeline consumes them consistently.
Write a short taxonomy before labeling the scene. For each class, state what belongs in it, what does not and how ambiguous parts should be handled. Give the taxonomy a version so later changes can be traced.
Distinguish class from instance
A class groups objects such as cabinet doors. An instance identifies an individual object. Decide which distinction your output needs and verify the behavior in your renderer or data-generation system.
NVIDIA's Replicator getting-started examples demonstrate semantic annotations within a configured data workflow. Use APIs appropriate to your installed version. Arbitrary custom attributes or a naming convention do not automatically become supported segmentation annotations.
Review the hierarchy and rendered output
Check labels on the composed scene, including referenced content and repeated objects. Then inspect a small annotation batch. The output is the test: a label visible in a property panel may not have the expected interpretation in your chosen annotator.
Our recommended review includes:
- A scene-wide view to find unlabeled or misclassified regions.
- Close views of interaction targets such as handles and drawers.
- Occluded objects and repeated instances.
- Reset and sequence behavior where identity continuity matters.
- The mapping from exported IDs to the names used in training.
Keep screenshots or small test frames with the review record. Correct the taxonomy or implementation before producing a larger dataset.
Treat completeness as a scoped claim
A scene can be labeled at object level without labeling every component. State the available granularity and any omissions. Do not describe annotations as complete or noise-free without defining and checking what that means for the task.
Inspect available metadata in the library. Read the ground-truth guide when selecting output modalities, or scope custom scene authoring when the labels and interaction parts need to match a particular facility.