Kookmin University team develops technology to optimize 3D object detection models for self-driving

Cho Hyun-joon, a master’s student at Kookmin University’s Graduate School of AI and SW, presents his team’s study conducted in collaboration with Hyundai Motor at the 2026 Design Automation Conference in Long Beach, California, from July 26-29. Courtesy of Kookmin University

A Kookmin University research team has developed SharedKD, a technology that efficiently compresses 3D object detection models for autonomous driving, in collaboration with Hyundai Motor.

The university said Tuesday that the findings were presented at the 2026 Design Automation Conference, a prestigious international conference in the field of design automation, held in Long Beach, California, from July 26-29.

Kim Jang-ho, a professor at the university’s College of Computer Science, led the research team. Two master’s students at the Graduate School of AI and SW, Cho Hyun-joon and An Sang-ho, also participated in the study and contributed to the development of the technology.

The university noted that the SharedKD technology differs from conventional knowledge distillation methods that use separate teacher and student models.

Instead, SharedKD uses the entire network of a single 3D object detection model as the teacher model and a subnetwork generated through pruning as the student model.

During training, SharedKD dynamically selects important structures based on gradients, enabling it to identify an efficient lightweight model while maintaining high accuracy.

The university said the significance of the study lies in moving beyond conventional approaches that reduce the size of pretrained artificial intelligence (AI) models after training.

Instead, the new technology enables the simultaneous exploration and training of lightweight models suitable for real-world deployment within a large-scale model. In addition, the overall model and the lightweight model are designed to support each other’s learning, enabling more efficient model exploration and training.

In autonomous driving applications, where both high perception accuracy and real-time processing are required, the technology can significantly reduce the computational workload and inference time of 3D perception models. It is therefore expected to have applications in automotive AI and edge AI systems.

The research also aligns with Kookmin University’s future-focused specialization strategy under “KMU Vision 2035: EDGE.” EDGE is short for entrepreneurship, digital transformation, global competency and ESG (environmental, social and corporate governance).

The university said the study is significant in that it demonstrates the university’s specialization strategy at the intersection of AI+X and mobility, two of its eight specialized fields, by applying AI model optimization technology to autonomous driving and conducting joint research with industry.

Meanwhile, the Design Automation Conference highlights a wide range of research advances, spanning semiconductor design, AI and machine learning systems and efficient AI implementation.