[Re] Pure Noise to the Rescue of Insufficient Data

TL;DR

This paper investigates using pure noise for data augmentation in imbalanced image classification, demonstrating that adding noise images improves minority-class accuracy and generalization. The proposed batch normalization method, which separately normalizes noise and natural images, enhances performance across varying augmentation levels and even benefits balanced datasets.

Why It Matters

It offers a simple yet effective strategy to address data imbalance without relying on complex augmentation or model architecture changes, making it practical for real-world scenarios with limited data. The method's robustness across different augmentation settings increases its applicability in diverse computer vision tasks.

Key Findings

• Oversampling with pure noise improves minority-class accuracy and generalization in imbalanced datasets.
• The proposed batch normalization layer, which uses affine parameters from natural images to normalize noise, outperforms standard baselines.
• Noise augmentation provides benefits even on balanced datasets and is robust to different levels of data augmentation.