日本語 
Dynamic Map Platform Co., Ltd. has released a dataset targeting intersections as a sample of data designed for use with AI on “Hugging Face,” a platform for the machine learning community where AI developers around the world share models and datasets.
This data is multimodal data that integrates point cloud data, camera images, high-precision location information, high-precision 3D map data, and 3D Gaussian Splatting (3DGS) data, and includes information on areas with a high risk of accidents.
This initiative is a concrete step toward realizing our data set business for physical AI, and aims to fully launch our AI data provision service, “Data for AI.”
Also Read: Murata and Synopsys Advance Engineering Simulation
Through the development and provision of high-precision 3D data, our company has been working to build a data infrastructure that reflects the real world in various fields, including autonomous driving. In addition, we have been developing “Data for AI” and accumulating knowledge about the ideal form of data for AI development.
In recent years, with the advancement of AI technology, the use of AI that targets the real world (physical AI) has expanded, and the importance of highly accurate data designed for specific applications has increased even further.
Against this backdrop, we will leverage our accumulated expertise to further develop “Data for AI” and promote the development and provision of AI-native data.
The AI-native data that our company aims to provide refers to high-precision 3D data that can be used as training data in AI model development and as digital twin data in simulations, and is designed and optimized to be immediately usable on the platform being used.
Typically, autonomous driving systems are developed using AI training based on real-world driving data such as dashcam footage. However, collecting large amounts of data and capturing rare scenes is time-consuming and costly. Against this backdrop, the importance of utilizing virtual environment data is increasing.
The sample data released this time is a multimodal AI-native dataset that utilizes the high-precision 3D data assets that our company has accumulated over many years. Point cloud data, multi-viewpoint camera images, high-precision location information, trajectory data, high-precision 3D map data, semantic annotation (data with semantic information), and 3D Gaussian Splatting (3DGS) data are provided in a temporally and spatially consistent manner.
This enables a learning and evaluation environment that accurately reproduces the real world. Furthermore, the feature information included in the map data can be used as annotation, contributing to the advancement of AI-based spatial recognition and the reduction of the Sim2Real gap (the difference between the real and virtual environments). In addition, because 3DGS data allows for the construction of a digital twin with a level of reproducibility close to that of the real environment, it can be used consistently from AI model training to the evaluation and verification of AI systems through simulation.
This sample targets real-world urban intersections with a high accident risk and can be used for advanced scene understanding and safety verification. It is particularly useful for the development of autonomous driving systems, and is expected to be used in a wide range of AI applications across various industries, including infrastructure management, urban development, traffic flow analysis, and disaster prevention and response.
This sample release is the first in a series, and we plan to expand our data lineup and continue releasing more data in the future, while also working on data design and development with a view to commercialization.
Our company will promote the development and provision of AI-native data, centered on “Data for AI,” and work towards advanced data utilization in preparation for the era of physical AI.
SOURCE: PRTimes
English