Enhancing Autonomous Vehicle Vision... Utilizing the "Vanishing Point" Technique of Renaissance Painters

Professor Kyungdon Joo’s Team at UNIST Develops Neural Network Model Incorporating Perspective Using Vanishing Point
Low-Cost Camera-Based Autonomous Driving and Robotics Applications Anticipated, Paper Accepted at IROS 2025

A new artificial intelligence technology has been developed to enable camera-based autonomous vehicles to perceive their surroundings more accurately.

This technology utilizes the "vanishing point," a geometric device that adds perspective to images.

On October 15, Professor Kyungdon Joo’s team at the Graduate School of Artificial Intelligence at UNIST announced that they have developed an artificial intelligence model called "VPOcc," which compensates for perspective distortion in information input through cameras.

Research team (from left) Professor Kyungdon Joo, Researcher Junsu Kim (first author). Provided by UNIST

The artificial intelligence systems of autonomous vehicles and robots recognize their surroundings using cameras or LiDAR sensors. Cameras are more affordable and lighter than LiDAR, and they provide rich information such as color and shape. However, since cameras represent three-dimensional spaces as two-dimensional images, there is significant distortion in size depending on distance.

Objects that are closer appear larger, while those farther away appear smaller, which can result in missing distant objects or overemphasizing nearby areas.

The research team addressed this issue by designing the artificial intelligence to reconstruct information based on the vanishing point. The vanishing point is a perspective technique established by Renaissance-era painters, referring to the point where parallel lines, such as road lanes or railway tracks, appear to converge in the distance. Just as people perceive depth on a flat surface by observing the vanishing point, the developed artificial intelligence model uses the vanishing point as a reference to more accurately restore depth and distance in camera images.

This model consists of three main modules: the VPZoomer module, which corrects images based on the vanishing point to reduce perspective distortion; the VPCA module, which extracts balanced information from both distant and close areas; and the SVF module, which combines the original and corrected images to complement each other's weaknesses.

Experimental results showed that VPOcc outperformed existing models in both spatial understanding capability (mIoU) and restoration performance (IoU) across various benchmarks. In particular, it was able to clearly predict distant objects and more accurately distinguish overlapping objects in road environments, which are crucial for autonomous driving.

This research was led by Researcher Junsu Kim at UNIST as the first author, with participation from Researcher Junhee Lee (UNIST) and researchers from Carnegie Mellon University in the United States.

Researcher Junsu Kim explained, "I started this research believing that incorporating the way humans perceive space into artificial intelligence would enable more effective understanding of three-dimensional space," and added, "This achievement maximizes the utility of camera sensors, which are more cost-effective and lightweight compared to LiDAR sensors."

Professor Kyungdon Joo expressed his expectations, saying, "The developed technology can be applied not only to robots and autonomous driving systems, but also to various fields such as augmented reality (AR) map creation."

Comparison of prediction results between the developed artificial intelligence model and the existing model.

The research results received the Silver Award at the 31st Samsung HumanTech Paper Award last March, and have been accepted for presentation at IROS (International Conference on Intelligent Robots and Systems) 2025, a prestigious conference in the field of intelligent robotics. This year’s conference will be held in Hangzhou, China, from October 19 to 25.

This research was supported by the Ministry of Science and ICT and the National Research Foundation of Korea.

(Paper title: VPOcc: Exploiting Vanishing Point for 3D Semantic Occupancy Prediction)

This content was produced with the assistance of AI translation services.