Solution for Large-scale Long-tailed Recognition with Noisy Labels
This addresses fine-grained image recognition for e-commerce, but it is incremental as it combines existing techniques without introducing new methods.
The authors tackled large-scale, long-tailed product recognition with noisy labels by using iterative data cleaning, classifier weight normalization, high-resolution fine-tuning, and test-time augmentation with ensemble models, achieving a 6.4365% mean class error rate.
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.