Bag of Tricks for Long-Tailed Visual Recognition with Deep Convolutional Neural Networks
[AAAI 2022] Bag of Tricks Github
文章目录
Bag of tricks:整理了四种针对长尾视觉识别任务的技巧,无太多新的东西。
Reweighing
- Cost-sensitive softmax cross-entropy loss (CS CE)
- Focal loss
- Class-balanced loss
Resampling
- Class-balanced sampling
- Square-root sampling
- Progressively-balanced sampling
Mix-up training
data augmentation Mix-up本身是一种data augmentation trick,但是和resampling结合对long-tailed recognition很有效,所以单列一项.
Input mixup
Manifold mixup
Fine-tuning after mixup training
Two stage
1在长尾数据集上跑朴素baseline不用resamlping不用reweight;
2在rebalanced数据上微调。
问题
- besides the extreme imbalance, the iNaturalist datasets also face the fine-grained problem
- Class-balanced loss (Cui et al. 2019) considers the real volumes of different classes, named effective numbers, rather than the nominal numbers of images provided by datasets.