Covering the food & beverage industry, beauty, fashion, retail markets, the Ministry of Land, Infrastructure and Transport, the Ministry of Agriculture, Food and Rural Affairs and related people and entities worldwide
UNIST professor enables AI to study summarized 3D data
Professor Shim Jae-young, left, of the Ulsan National Institute of Science and Technology (UNIST), and his research team have developed a new method that enables artificial intelligence to analyze 3D data in a summarized form, reducing both time and cost for 3D data analysis. Courtesy of UNIST
By Ko Dong-hwan
A professor at the Ulsan National Institute of Science and Technology (UNIST) has developed a new technology that allows artificial intelligence (AI) models to study 3D data in a “summarized” form while still achieving a high level of accuracy, reducing both time and cost in 3D data analysis.
Shim Jae-young and his team at the UNIST AI Graduate School in Ulsan said Monday that they have developed a “dataset distillation” technology for large-scale 3D data. The technology extracts the core of a given dataset and compares it with the original to retain completeness, enabling effective compression of a 3D point cloud — data composed of randomly distributed dots representing an object — without sacrificing accuracy during analysis.
Shim developed the technology after finding that summarizing a 3D point cloud for AI training is particularly difficult. Because the dots in a point cloud are arranged in a random, orderless manner and often represent objects rotated at various angles, AI systems struggle to extract a reliable 3D summary and frequently mismatch it with the original dataset. Shim said these characteristics were “critical hurdles” for AI systems working with 3D data.
To overcome these challenges, Shim introduced a dataset distillation method that automatically rearranges the dots according to their “meanings” and optimizes rotational angles through a “learnable rotation” process, enabling AI models to analyze them more effectively.
In tests using the new technology, Shim found that AI models analyzing a 3D dataset compressed to one-twenty-fifth of its original size achieved an accuracy rate of 80.1 percent. This closely approached the 87.8 percent accuracy the same models recorded when trained on the full dataset.
“This technology will provide a fundamental solution to the mismatch between original and summarized 3D data caused by random point arrangements and rotations,” Shim said. “It will greatly benefit sectors such as autonomous driving, drones, robotics, and digital twins, where large-scale 3D data is essential.”
Shim’s research was funded by the Ministry of Science and ICT and the Institute of Information & Communications Technology Planning & Evaluation.
His paper, "Dataset Distillation of 3D Point Clouds via Distribution Matching," has been selected for presentation at the Conference on Neural Information Processing Systems, which will be held from Tuesday to Sunday in San Diego.