Solution for Large-scale Long-tailed Recognition with Noisy Labels

TL;DR

This paper presents a comprehensive solution for large-scale, long-tailed commodity image recognition with noisy labels, utilizing advanced architectures and data cleaning techniques to improve accuracy in an ecommerce challenge.

Contribution

It introduces a novel combination of model architectures and data processing strategies specifically designed for noisy, imbalanced, and fine-grained recognition tasks.

Findings

Achieved 6.4365% mean class error rate in the competition

Identified key techniques: data cleaning, classifier normalization, high-res finetuning, test augmentation

Demonstrated effectiveness of CNN and Transformer models in large-scale recognition

Abstract

This is a technical report for CVPR 2021 AliProducts Challenge. AliProducts Challenge is a competition proposed for studying the large-scale and fine-grained commodity image recognition problem encountered by worldleading ecommerce companies. The large-scale product recognition simultaneously meets the challenge of noisy annotations, imbalanced (long-tailed) data distribution and fine-grained classification. In our solution, we adopt stateof-the-art model architectures of both CNNs and Transformer, including ResNeSt, EfficientNetV2, and DeiT. We found that iterative data cleaning, classifier weight normalization, high-resolution finetuning, and test time augmentation are key components to improve the performance of training with the noisy and imbalanced dataset. Finally, we obtain 6.4365% mean class error rate in the leaderboard with our ensemble model.