Research Team Led by Professor PARK Ju-hyun Proposes Novel Method for Estimating Geometric Models in Computer Vision

Strong Academic and Industrial Interest for Applications in Autonomous Driving, 3D Reconstruction, Robotics Control, and Augmented Reality (AR)

First-Author Paper by Dr. CHOI Yeon-gyu Published in IEEE TPAMI, World’s Leading Journal in Computer Vision

[April 27, 2026]

<Researcher CHOI Yeon-gyu from the Department of Electrical Engineering at YU>

A research team led by Professor PARK Ju-hyun of the Department of Electrical Engineering at YU (President CHOI Oe-chool), under the Nonlinear Dynamics Laboratory, has developed an advanced optimization technology for geometric modeling in computer vision applicable to autonomous driving and robotics, drawing significant attention from both academia and industry.

 This study was conducted with Dr. CHOI Yeon-gyu of YU’s Department of Electrical Engineering as the first author. The research paper is scheduled to be published in the May 2026 issue of IEEE Transactions on Pattern Analysis and Machine Intelligence (IEEE TPAMI) [JCR top 0.7%, Impact Factor 18.6], the world’s most prestigious journal in artificial intelligence and computer vision.

 Dr. CHOI explained, “This study focuses on advancing techniques that precisely model the geometric relationship between two images captured from different viewpoints of the same 3D scene by two cameras in computer vision. Conventional approaches have structural limitations in deriving optimal models in environments with significant noise or outliers. In this work, we address these limitations by introducing a novel post-processing mechanism that maximizes the inherent performance of the algorithm.”

The newly proposed method, termed MEPC (Multi-Estimation-based Parameter Centroid), goes beyond the conventional practice of selecting a single model with the highest statistical score. Instead, it determines a central set of model parameters based on multiple hypothesis candidates generated through repeated estimation processes. This approach reduces the impact of data distortion and enables more accurate and stable geometric modeling.

 Notably, this technology can be applied across a wide range of vision-based industries that require highly precise geometric model estimation, including autonomous driving, 3D reconstruction, robotic control, and augmented reality (AR). In these fields, minimizing the effects of data distortion while accurately capturing the geometric structure of real-world environments is critically important.

 The research team stated, “This achievement expands the limits of how precisely model estimation can be optimized even in noisy environments. It is expected to significantly enhance the technological competitiveness of systems requiring high-precision sensor pose estimation, such as multi-sensor fusion, 3D Gaussian Splatting (3DGS), and Simultaneous Localization and Mapping (SLAM).”

 Building on this outcome, the team plans to conduct follow-up research aimed at advancing next-generation multimodal AI convergence and intelligent mobility systems. This research was supported by the National Research Foundation of Korea.