Class imbalance is a common issue in many real-world problems. Instead of existing solutions like resampling or reweighting, the authors propose augmenting minority classes with oversampled images and pure noise images, showing state-of-the-art results on long-tail (LT) datasets. The authors speculate that this method improves performance since pure noise images add both magnitude and stochasticity to the gradients, whereas oversampling only increases gradient magnitude without altering the gradient direction.
Because pure noise images are from a different distribution, the authors also devise a new batch normalization scheme called Distribution-Aware Routing Batch Normalization (DAR-BN), where a separate batch normalization is used for real images and pure noise images.
The authors performed ablation studies on DAR-BN and pure-noise images. Check the paper to learn more!