The research team of Professor Jang Bu-ru, Department of Artificial Intelligence,

Paper Accepted at ECCV 2024, the World's Foremost Computer Vision Conference

▲ Professor Jang Bu-ru, Department of Artificial Intelligence

A paper with a Seoul National University research team, on which Professor Jang Bu-ru of the Department of Artificial Intelligence served as corresponding author, has been accepted at the European Conference on Computer Vision (ECCV 2024, BK-recognized IF=2), one of the world's three leading computer vision conferences.

The team will present a paper titled 'Exploiting Semantic Reconstruction to Mitigate Hallucinations in Vision-Language Models' at ECCV 2024. The research releases ESREAL, a fully unsupervised learning method for mitigating the hallucination problem in large vision-language models.

Without any hallucination-labelled data, ESREAL uses text-to-image generation technology to reconstruct an image from hallucinated text generated for a given image, and detects where and what kind of hallucination occurred by comparing that reconstruction with the original image.

Professor Jang Bu-ru said: "This research can mitigate the hallucination problem in large vision-language models, and through that we expect ultimately to be able to improve the reliability of AI models. This result is all the more noteworthy given that every author is an undergraduate."

ECCV 2024 will be held in Milan, Italy, from 29 September to 4 October.

▶ Paper title: Exploiting Semantic Reconstruction to Mitigate Hallucinations in Vision-Language Models

▶ Paper link: https://arxiv.org/abs/2403.16167

Source: Sogang Story – News – Research Achievements https://sogang.ac.kr/ko/story/media-center?tab=3