3rd Place Solution to ICCV LargeFineFoodAI Retrieval
Yang Zhong, Zhiming Wang, Zhaoyang Li, Jinyu Ma, Xiang Li
TL;DR
The paper tackles fine-grained food image retrieval under disjoint training and testing classes by presenting a multi-backbone framework trained with a hybrid ArcFace plus Circle loss and enhanced with test-time augmentation, model distillation, and a diffusion-based plus k-reciprocal reranking post-processing pipeline. The approach combines discriminative embedding learning with a robust ensemble and a two-stage reranking strategy, achieving mAP@100 of $0.81219$ on the public leaderboard and $0.81191$ on the private leaderboard. Key technical contributions include the loss fusion formula $L = δ_{0}L_{a} + δ_{1}L_{c}$ with $δ_{0}=1$ and $δ_{1}=1/β$, ArcFace parameters margin $0.2$ and scale $32$, and Circle loss with $m=0.25$ and delta $32$, as well as diffusion-based ($S$) and k-reciprocal ($D$) reranking with $S_{final}=S - λD$. The work demonstrates a practical, high-accuracy retrieval pipeline for large-scale fine-grained visual search, highlighting the effectiveness of combining discriminative losses, distillation, TTA, and diffusion-based reranking for transfer under non-overlapping class splits.
Abstract
This paper introduces the 3rd place solution to the ICCV LargeFineFoodAI Retrieval Competition on Kaggle. Four basic models are independently trained with the weighted sum of ArcFace and Circle loss, then TTA and Ensemble are successively applied to improve feature representation ability. In addition, a new reranking method for retrieval is proposed based on diffusion and k-reciprocal reranking. Finally, our method scored 0.81219 and 0.81191 mAP@100 on the public and private leaderboard, respectively.