Reproduce “$S^{2}$-FracMix: Label-Preserving Self-Saliency Mixup Augmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Khawar Islam, Arif Mahmood, Xin Jin, Naveed Akhtar
The S2-FracMix framework introduces a data augmentation technique that enhances the generalization and robustness of deep visual models through a combination of single-image self-saliency analysis, targeted fractal injection, and a multi-mode mixing strategy. This approach consistently achieved state-of-the-art accuracy across general and fine-grained classification tasks, transfer learning, and improved robustness against corruptions and adversarial attacks while maintaining computational efficiency.
Key insights
- Augmenting a single image based on its self-saliency effectively preserves label consistency and semantic integrity, avoiding inconsistencies introduced by mixing disparate images.
- Localized injection of fractal patterns into salient regions compels models to learn features robust to high-frequency perturbations within discriminative areas, while maintaining overall image context.
- A stochastic multi-mode mixing strategy, combining self-saliency-guided fractal augmentation with other efficient methods, diversifies training inputs, leading to more robust object representation learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.