□ A research team led by Soon Kwon and Jinhee Lee from the Division of Automotive Technology at DGIST (President Kunwoo Lee) has successfully developed “MultipleTeachers,” a 3D semi-supervised object detection (SSOD) framework that achieves state-of-the-art (SOTA) performance in environments with extremely few labels.
□ This technology introduces a new learning strategy that groups similar objects together to form teacher networks for each category, which cooperatively generate pseudo labels. In addition, the “PointGen” module is integrated to address the issue of rare LiDAR points, thereby significantly improving the recognition accuracy of key objects, such as automobiles, pedestrians, and motorcycles. A new paradigm has been suggested where it contributes to the advancement of safety-oriented autonomous driving cognitive technology by inducing effective learning from limited data.
□ Advancements in autonomous driving technology have often been hampered by the high cost and time required for labor-intensive data labeling. The research team at DGIST developed semi-supervised and self-supervised learning methods that effectively combine a small amount of labeled data with a large volume of unlabeled data. This approach significantly reduces dependency on labeled data and has demonstrated outstanding performance in experimental settings.
□ Furthermore, in collaboration with the DGIST startup FutureDrive, the team has created a custom LiDAR dataset, LiO, that accurately reflects the urban environments in Korea. LiO guarantees high quality through expert reviews conducted at least three times for seven classes, based on data collected from one 128-channel LiDAR sensor and six cameras. It can be utilized in a wide range of experiment settings, since it consists of 35.8 objects on average, along with nearly 21,000 and 96,000 labeled and unlabeled frames, respectively.
□ The performance verification also produced superior results. The proposed technology outperformed conventional SOTA technologies by recording a mAP of 47.5 in the Waymo Open Dataset (1% labels), 72.2 in the KITTI dataset (2% labels), and 61.4 in the LiO Large (15% labels) dataset. In particular, the recognition performance was noticeably improved for small objects such as pedestrians and motorcycles, thus contributing to safety enhancement in urban areas.
□ Dr. Lee stated, “It is an honor to present the cognitive technology of DGIST at ICCV 2025, one of the world’s best vision conferences,” and added “we shall publicly release the LiO dataset to share knowledge with the research communities and expand its application across diverse fields, such as autonomous driving, smart cities, and logistics robotics.”
□ This research was supported as a part of the DGIST institutional project and the Ministry of Science and ICT’s R&D Special Zone Promotion project. The research findings shall be officially announced at ICCV 2025, scheduled to be held in October.
