Fine-Grained Image Retrieval with Neighbor-Attention Label Correction

Shunan Mao, Shiliang Zhang

Keywords: noisy label, meta learning, fine-grained image retrieval

TL;DR: A Neighbor-Attention Label Correction model trained by nested optimization is proposed to correct noisy label in fine-grained image retrieval

Abstract: This paper studies noise-resistant deep model training for the fine-grained image retrieval task, which has a unconstrained target label space and suffers from the difficulty of acquiring accurate fine-grained labels. A Neighbor-Attention Label Correction (NALC) model is proposed based on the meta-learning framework to correct labels during the training stage. A training batch and a validation batch are sampled from the training set, which hence allows to optimize the NALC model by referring to the validation batch. We also propose a novel nested optimization for the meta-learning framework to enhance the optimization efficiency. The training procedure consistently boosts the label accuracy in training batch, which in-turn ensures a more accurate training set. Experiments results show that our method boosts the label accuracy from 70% to 97+% and it outperforms recent works up to 11.5% in rank1 accuracy on various fine-grained image retrieval tasks, e.g., fine-grained instance retrieval on CUB200 and CARS, as well as person re-identification, respectively. Ablation studies also show the NALC generalizes well on different types of noises, e.g., Asymmetric, Pair-Flip, Pattern noises, etc.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Applications (eg, speech processing, computer vision, NLP)