Image processing AI startup MayI (CEO Park Jun-hyeok) announced on the 6th that its paper titled 'A Study on Camera Bias in Human Re-Identification AI Model' was 'adopted' by the 'International Conference on Representation Learning (ICLR) 2025', the world's most prestigious AI academic conference, proving the technological competitiveness of its solution 'mAsh'.
ICLR is a leading academic conference in AI research in the world, and is considered one of the top three global AI conferences along with NeurIPS and the International Conference on Machine Learning (ICML). Every year, leading research institutes and companies from around the world participate, share the latest research trends through various sessions, and select excellent papers through strict review. The paper submitted by MayI was selected as the 'Spotlight', which is given to the top 5%, and was recognized for its innovation.
In this paper, 'Exploring the Camera Bias of Person Re-identification (Song Myeong-seo, Park Jin-woo, Lee Jong-seok)', MayEye analyzed the limitations of existing AI models that have difficulty clearly identifying the same person due to environmental differences in each CCTV, and studied methods to alleviate the bias between cameras so that consistent accuracy can be maintained even under various conditions.
Based on this, MayI proposed a ‘normalization technique to reduce camera bias’ and improved the ‘learning method of unsupervised human re-identification technology.’ Since the AI model correctly recognizes people even in new spaces that it has not learned, and implements high accuracy with a small number of personnel, it can be applied to various stores to maximize operational efficiency.
In particular, MayI's self-developed re-identification (Re-ID) technology, including the technology introduced in this paper, has recorded an accuracy of 92% in internal tests, which outperforms the latest academic models (ISR, 66%), and is being applied to mesh to precisely analyze visitor data in various offline spaces.
“We are delighted that our research to improve the accuracy of mesh has been recognized worldwide,” said Song Myeong-seo, lead research team member and first author of the paper. “We expect that this will allow us to provide more customers with reliable analysis results.”
Kim Chan-gyu, CPO of May-I, said, “The human re-identification model presented in this paper is May-I’s core technology that provides precise store visitor data while protecting personal information,” and added, “We will continue to strive to provide more accurate data with world-class technology.”
Meanwhile, MayI acquired a patent for AI-based image analysis technology from the U.S. Patent and Trademark Office in September of last year. Through this, it is strengthening the legal protection of technology that analyzes customer behavior in offline stores in real time and solidifying its position in the global market.
- See more related articles
You must be logged in to post a comment.