[Re] CUDA: Curriculum of Data Augmentation for Long‐tailed Recognition

摘要

In this reproducibility study, we present our results and experience during replicating the paper, titled CUDA: Curriculum of Data Augmentation for Long-Tailed Recognition(Ahn et al., 2023).Traditional datasets used in image recognition, such as ImageNet, are often synthetically balanced, meaning each class has an equal number of samples. In practical scenarios, datasets frequently exhibit significant class imbalances, with certain classes having a disproportionately larger number of samples compared to others. This discrepancy poses a challenge for traditional image recognition models, as they tend to favor classes with larger sample sizes, leading to poor performance on minority classes. CUDA proposes a class-wise data augmentation technique which can be used over any existing model to improve the accuracy for LTR: Long Tailed Recognition. We successfully replicated all of the results pertaining to the long-tailed CIFAR-100-LT dataset and extended our analysis to provide deeper insights into how CUDA efficiently tackles class imbalance. The code and the readings are available in https://anonymous.4open.science/r/CUDA-org--C2FD/README.md