ML data labeling standard for ADAS, AV development
Deepen AI led the working group within ASAM to launch OpenLABEL – the first-ever global standard for multi-sensor data labeling and scenario tagging. Data labeling is fundamental to teach machines to perceive and understand the world, and open-labeled datasets have fueled the growth and development of machine learning (ML) methods for perception in ADAS and autonomous vehicles (AVs) over the last decade (Waymo open dataset, KITTI, NuScenes, Argoverse, etc.).
However, a major problem, say the organizations, is that all of these datasets – and the majority of the organizations building labeled datasets – use their own taxonomies, formats, and data models to encode the information in a static, goal-specific, and heterogeneous way, producing the following problems:
- Limited reuse of annotated datasets.
- Challenges regarding maintenance and updating of the annotations
- Limited sharing of datasets across the industry and between industry and academia.
- Negative impact on the quality of annotations.
OpenLABEL is designed to solve these issues by providing a common format and data model to structure and organize the information regarding multi-sensor data labeling (cameras, lidar, etc) in a standardized manner. OpenLABEL also defines a standardized set of tags and the data model to categorize and organize test scenarios for ADAS and AVs.
“ASAM OpenLABEL is the first-of-its-kind standard, and definitely not a conventional one,” says Nicola Croce, Technical Program Manager, Deepen AI, OpenLABEL project leader and member of the ASAM Technical Steering Committee. “It is underpinned by an entirely novel approach to labeling that enables a much more efficient way to manage and maintain labels and their semantics through the use of ontologies. We are confident that OpenLABEL will significantly help the industry in iterating faster, sharing data more efficiently, and ultimately deploy safe ADAS and AV systems sooner.”
Peter Voss, Managing Director ASAM e.V says, “ASAM OpenLABEL is a completely new standard that unifies multisensor dataset labeling and scenario tagging and facilitates their exchangeability. The standardized format and common data models as well as the application of ontologies as a basis, will lead to a noticeable increase in the quality of the labeled datasets and thus in the safety of autonomous driving. Especially if used in combination with ASAM’s OpenXOntology (to be released in December 2021), ASAM OpenLABEL will lead to more efficient development cycles, freeing up capacities for other tasks. As this is worldwide the first standard to tackle the issue of labeling, we are sure that ASAM OpenLABEL will soon be a widely used standard in the industry.”
OpenLABEL, say the organizations, will significantly contribute to increasing the quality of annotated datasets, their maintainability, the ability to share, combine and repurpose them.
For more, see the webinar: Introduction to ASAM OpenLABEL V1.0.0
Related articles:
Autonomous driving dataset is free to researchers
AI startup looks to ‘democratize access’ to intelligent data
AV safety collaboration targets complex AI reliability challenges
Whitepaper: Building smart IoT devices with automated ML
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
