ASAM OpenLABEL V1.0.0
on: June 09, 2021 | 1:00 pm - 4:00 pm CEST
via: Teams Meeting
This webinar will introduce the ASAM OpenLABEL which is currently under development and provide a status update on the expected outcome. We ask all experts and potential users of this standard to take part and to provide feedback on the current scope and potential further requirements. standard
Please note that this webinar is intended to be an active event. Our goal is to develop a that joins the ranks of the widely used and widely adopted ASAM OpenX standards. For that reason, the webinar is not just about presenting the standard, but also about collecting your feedback, ideas and further requirements. We invite you to make use of this opportunity: Your feedback will help to shape ASAM OpenLABEL so that you can get the most out of it. standard
About ASAM OpenLABEL
The ASAM OpenLABEL standardization project aims at standardizing an annotation format and methods for labeling multisensor data streams which are used for training and validating machine learning models that are usually part of an automated driving system perception stack. ASAM OpenLABEL will provide a guideline on how the labeling methods and definitions should be used. The project group is collaborating closely with the ASAM OpenXOntology project.
Why is a standard needed?
From working with different customers, a significant fragmentation emerged in the way each individual organization categorizes and describes the objects populating the driving environment. Such categorizations and descriptions are the fundamental building block of any Autonomous Driving System’s (ADS) perception stack, since it is through them that an ADS comes to a basic and profound understanding of the status of around its surrounding.
The lack of a common labeling in the industry is the root cause of several different issues: standard
- Hampered Vehicle2Vehicle Interaction: The different descriptions and understandings of surroundings may cause casualties in complex situations involving two or more different ADSs
- Precluded sharing: It is a highly difficult if not impossible task to share data across organizations that adopted different labeling taxonomies and specifications
- Reduced annotation quality: Each individual labeling task requires ad-hoc training and even development of custom software functions that translate into a higher probability of errors and thus a threat to safety
- Deprecation of old labels: Long-term operation of ADS development imply changes in quantity and comprehensiveness of labels to be produced considering the evolution of the driving scenes, new sensors, and scenarios. As a consequence, a flexible descriptive language is required to absorb future extensions and modifications of labels and guarantee backward-compatibility
How will ASAM OpenLABEL solve these challenges?
The OpenLABEL particularly focusses on the following topics: standard
- Objects and Scene Labeling: The standard shall describe how single objects can be labelled and how these labeled objects can be labeled in the context of a scene. This work package has several perspectives: conditional labels, event labels, action labels, relation labels.
- Coordinate systems and sensor streams synchronization, metadata etc.
- Scenario Tagging: The standard shall define scenario labels on a meta level. This includes labels that can be derived from the content of the scenario as well as labels which are non-derivable.
- Label file format for multisensor data: The standard shall provide a JSON schema based on the input given in the concept paper. The aim is to enrich scenario artifacts (provided in any Scenario Description Language or other representations) with a set of tags to fulfil the use case of organizing, searching and filtering scenarios in scenario databases - such as the www.safeypool.ai Scenario database.
Important Features of ASAM OpenLABEL
ASAM OpenLABEL will be represented in a JSON format and can therefore be easily parsed by tools and applications. ASAM OpenLABEL will specify which coordinate systems are used as reference for the . This already facilitates the conversion a lot. label
Extended Labeling Objects
ASAM OpenLABEL will also provide methods to objects in a scene (one point in time/ frame) as well as across multiple scenes by enhancing the methods to label actions, intentions and relations between objects. label
Labeling Different Data Types
The ASAM OpenLABEL format will be capable of managing different types of labeling methods, for different types of data. This includes 2D and 3D bounding boxes, the rotation of 3D bounding boxes, semantic segmentation of images and point clouds. These semantic segmentations can be either instance classes, single/multi-class, partial or full classes.
It is important that the labeling fits into the taxonomy definitions of a user/company. For that reason, the project group intends to provide ASAM OpenLABEL with the ability to import ontologies and taxonomies for the labeling process. The ASAM OpenLABEL project group is closely interacting with the ASAM OpenXOntology project to align ASAM OpenLABEL with the OpenX domain model and to provide requirements for the ASAM OpenXOntology . standard
ASAM OpenLABEL V1.0.0 is foreseen to be released in Nov 2021.