FOCUS AI Ranks First in International Retail Video Analytics AI Challenge

FOCUS AI announced on the 14th that ‘PASTEL’, a retail video analytics AI (artificial intelligence) model developed with the participation of its researchers, took first place in the Physical Retail AI Challenge held in connection with the international computer vision conference (WACV 2026).
The competition was affiliated with a WACV 2026 workshop, and participation and evaluation took place on Kaggle, an AI competition platform. In the ‘RetailAction Track 3’ category, PASTEL recorded an official evaluation score of 0.6776, ahead of the second-place team (0.5021).
WACV is a conference on computer vision applications sponsored by the IEEE Computer Society and the Computer Vision Foundation (CVF), and is an international conference in the same family as CVPR and ICCV. Kaggle is a data and AI competition platform acquired by Google in 2017, where models submitted by participating teams are scored on the same criteria and ranked.
PASTEL is a model jointly developed by researchers from FOCUS AI and SafeAI. It analyzes video from multiple cameras installed in a store together with customer pose information to identify simultaneously when and where interactions between customers and products occur.
Unlike existing retail video analytics, which has focused on determining whether a customer behavior occurred, PASTEL also estimates the time interval in which the behavior took place and the point of product contact in each camera view. It reflects store environments where parts of the body are occluded by shelves or other customers and the shooting angle differs for each camera.
The company presented its research results as a paper at the 17th International Conference on Ubiquitous and Future Networks (ICUFN 2026) held in Milan, Italy, from the 7th to the 10th. The paper is titled ‘PASTEL: Pose-Aware Spatio-Temporal Event Localization of Interactions in Multi-view Retail Videos’ and covers the technology and performance of analyzing interactions between customers and products in a multi-camera environment.
The research used about 21,000 video samples collected from 10 operating stores and a 41-hour Retail Action dataset. In the experiments, PASTEL recorded an overall spatio-temporal detection performance (mAP) of 73.3, a spatial localization performance of 78.3, and a temporal interval detection performance of 90.7. The overall performance of a model using only camera video was 57.7.
FOCUS AI plans to apply PASTEL to smart retail and intelligent video monitoring by linking it with existing CCTV, NVR, and video monitoring infrastructure. The company intends to expand store operation support functions such as product pick-up and return, shelf interactions, and customer behavior and anomaly analysis, and to broaden the scope of application to multi-camera video analytics in industrial and commercial facilities.
A FOCUS AI representative said, “Through this award and the presentation at the international conference, we confirmed our technological capability to analyze interactions between customers and products even in complex store environments,” and added, “While securing core video analytics technology and widening our technology cooperation network, we will advance smart retail and intelligent video monitoring solutions that combine AI with existing video security infrastructure.”
Meanwhile, FOCUS AI, founded in 2012, listed on the KOSDAQ market in 2021 through a merger with Eugene SPAC No. 5. Its main products are physical security solutions such as AI cameras, network video recorders (NVRs), and video monitoring platforms. At its annual general meeting of shareholders, the company changed its former name, FOCUS HNS, to its current name in March 2025.
