Semantic Labels#
Overview#
This capability provides requirements for semantic label attributes. These semantics may be used to provide ground truth for ML training by verifying perception system identification and classification of objects
Summary#
Geometry are semantically labeled
Semantic labels can be inherited from ancestral prims as well as bound materials
NVIDIA uses the QCode taxonomy for its Omniverse Asset libraries. Assets intended to be used alongside NVIDIA assets should use the same taxonomy.
Granularity#
At least one semantic label is should be provided per asset, to identify it as a “car,” “pedestrian,” “forklift,” etc. Additional semantic labels on parts within the asset (e.g. “tire,” “windshield,” “hubcap,” etc.) increases the usefulness of the semantic labels for ML training.
Taxonomy#
Wikidata Taxonomy#
NVIDIA chose to utilize Wikidata for Omniverse Asset libraries - an open-source taxonomy database with over 115 million searchable items including objects, brands, and locations.
Different/Multiple Taxonomies#
While NVIDIA utilizes Wikidata Q-codes as the primary taxonomy, developers may implement alternative taxonomies that better fit their specific use cases. Thanks to the OpenUSD SemanticsLabelsAPI’s “Multiple Apply” schema design, multiple taxonomies can coexist on the same object without conflict, allowing for flexible and comprehensive semantic labeling.
Schema / OpenUSD Specification#
Semantic Labels are defined with the SemanticLabelsAPI schema. Available in OpenUSD 24.11 and later.
USDA Sample#
def XForm "Vehicle" (
prepend apiSchemas = ["SemanticsLabelsAPI:wikidata_qcode"]
)
{
token[] semantics:labels:wikidata_qcode = ["Q42889"]
}
Requirements#
Tags |
Summary |
Compatibility |
Validator |
|---|---|---|---|
🔑 |
core usd |
||
✅ |
open usd |
||
✅ |
nvidia omniverse |
||
⛔ |
rtx |