LargeFineFoodAI: ICCV 2021 Food Vision Seminar and Challenge Overview

The ICCV 2021 LargeFineFoodAI seminar, jointly organized by Meituan Vision Intelligence, the Chinese Academy of Sciences Computing Institute, Beijing Zhiyuan, and the University of Barcelona, was held online on October 16 (19:00‑22:30 Beijing time).

Invited talks featured Ramesh Jain (University of California, Irvine) on building personalized food models for enjoyment and health, Kiyoharu Aizawa (University of Tokyo) presenting FoodLog Athl – a multimedia food‑recording platform for dietary guidance, and Petia Radeva (University of Barcelona) discussing how uncertainty modeling can enhance food recognition.

The associated challenge attracted 143 domestic and international teams (including Tsinghua, USTC, Nanjing Tech, Alibaba, OPPO, etc.) and comprised two tracks: large‑scale fine‑grained food recognition and large‑scale fine‑grained food retrieval. Top‑3 rankings for the recognition track were Joyy‑cv (92.058 %), NJUST‑PCALab (92.032 %), and OPPO Research Institute (92.019 %). For the retrieval track, the top‑3 were DeepBlueAI (82.813 % mAP), USTC‑NELSLIP (82.199 % mAP), and OPPO Research Institute (81.191 % mAP).

A new dataset released with the seminar contains over 1,000 fine‑grained food categories and more than 500,000 images (153‑1999 images per class), built with expert‑validated food ontology and designed to pose significant class‑imbalance challenges for recognition and retrieval.

Oral presentations included two accepted papers: “Fine‑Grain Prediction of Strawberry Freshness using Subsurface Scattering” (Carnegie Mellon University) and “Online Continual Learning for Visual Food Classification” (Purdue University).

Attendees could join the LargeFineFoodAI technical group by sending the keyword, receive the meeting link, and participate in a Q&A session with prize incentives.