OpenLABEL

OpenLABEL is an open annotation standard, published by the standards body ASAM, that defines a common way to describe labels on multi-sensor data and to tag scenarios. Its purpose is to make annotations portable, so that labels created in one tool or by one organization can be understood and used by another without bespoke conversion. It specifies a data model, expressed in JSON, for organizing labels across data from cameras, lidar, radar, and other sensors, and it covers a range of annotation types including 2D and 3D bounding boxes, the rotation of 3D boxes, and semantic segmentation of images and point clouds.

Beyond individual object labels, OpenLABEL also provides a way to categorize and tag scenarios, which connects it to the broader task of describing situations rather than just objects. It is frequently described as the first standard aimed specifically at labeling multi-sensor data and scenarios together. Because it uses JSON, its files are both machine-parseable and human-readable, and it supports referencing ontologies so that label names and categories have unambiguous, shared meaning. In a field where annotation formats have often been tool-specific, OpenLABEL represents an effort to establish a common language.

OpenLABEL specifies how labels and their associated information are structured and serialized, using a JSON schema so that files can be exchanged between tools and remain readable to both software and people. The data model organizes annotations, associates them with the relevant sensor data, and supports the geometric and semantic label types common in autonomous systems, from oriented 3D boxes to point-cloud segmentation. To keep the meaning of labels unambiguous, it allows the use of ontologies that define what each category and tag refers to. Scenario tagging extends the same idea from individual objects to whole situations, giving a standardized vocabulary for describing what a segment of data contains.

OpenLABEL matters because inconsistent, tool-specific annotation formats are a real obstacle to sharing and combining data across teams and vendors, and a common standard reduces that friction. For anyone curating or exchanging multi-sensor datasets, a portable format means labels do not have to be painstakingly translated every time data moves between systems, which lowers cost and reduces errors. As physical AI increasingly depends on pooling and standardizing heterogeneous data, shared annotation standards like OpenLABEL become part of the foundation that makes such pooling practical.

It supports a range of annotation types used in autonomous systems, including 2D and 3D bounding boxes, the rotation of 3D boxes, and semantic segmentation of both images and point clouds. It also provides a way to categorize and tag scenarios.

Because annotations are often created in tool-specific formats that others cannot easily use, a shared standard makes labels portable across tools and organizations. That reduces the need for custom conversion, lowering cost and the risk of errors when data is exchanged.

It uses a JSON-based data model, so its files are both machine-parseable and human-readable. It also supports referencing ontologies so that label names and scenario tags carry unambiguous, shared meaning.

Label propagation, 3D bounding box / cuboid, Label track, Scenario mining, Data curation.