Synthetic Data Generation Platform

The virtual data generation platform is a platform for artificial intelligence learning, especially for deep learning including image-based object detection. This platform virtually generates learning data. When these virtual learning data and a small amount of real data are combined, learning performance can be greatly increased. In particular, it generates synthetic data in adverse environments or various spatial conditions with several generation approaches such as 3D graphic, 2D GAN, Nerf, stable diffusion. It also generate synthetic IR data through stable diffusion approach with small paired CCD and IR camera data. To solve problems such as lack of data acquisition or time consuming, it quickly generate learning data and witness network model performance in advance.

Specifically, it automatically changes environmental conditions (rainfall, snowfall, fog), atmospheric conditions (lighting, dusk, etc.), and spatial conditions. It also divides the azimuth and elevation angles of the measured object automatically and randomizes the distance, etc. Therefore, it provides sufficient domain information necessary for learning. Even if used quickly or after securing sufficient actual data, there may be a lack of diversity and lack of proportionality in the data. It can be used to resolve data bias that occurs and maximize learning performance.

Currently, we are generating both CCD and IR images, and extending radar images for deep learning for autonomous vehicles. We are continuously focusing on the development of various sensor models that can be applied and produced. We pursue evolutionary development based on the automatically generated basic version for non-deformable objects (humans, animals, etc.) through applying state of the art technology.