Reproduce “InstructMixup: Instruction-Guided Salient Patch Editing for Robust Data 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
Researchers at The University of Melbourne, Information Technology University, and Westlake University developed InstructMixup, an augmentation method that edits salient image patches using instruction-guided generative models and fractal textures, all within a single image. This approach improves classification accuracy, robustness to corruption, and calibration across various vision tasks and architectures, while minimizing runtime computational overhead.
Key insights
- Augmenting within a single image while verifying label consistency for generative edits can effectively introduce appearance diversity without compromising semantic integrity or accuracy.
- Decomposing complex generative augmentation into an offline pre-computation stage and a runtime application of cached edits makes sophisticated techniques computationally efficient during model training.
- A principled approach combining structured transformations (generative edits, rotations) with localized additive perturbations (fractal injection) simultaneously enforces invariance and penalizes loss curvature in discriminative regions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.