Embedded Files
Data Augmentation
Data Augmentation
Data augmentation increases the size and diversity of training data by creating modified versions of existing samples. It improves generalization, reduces overfitting, and acts as an implicit regularizer. In computer vision, augmentation exposes models to broader appearance variations without requiring new labels.
We propose a dual-region strategy that improves robustness while reducing dependence on large labeled datasets. The method applies:
Noise perturbations targeted at foreground objects.
Spatial shuffling of background patches.
This structured augmentation yields stronger generalization across tasks without additional supervision.
Highlights
Separates foreground and background to apply task-aware perturbations.
Encourages invariance to background bias and spurious context.
Improves out-of-distribution robustness in recognition settings.
Shuffle PatchMix is an intra-image patch mixing approach that shuffles and blends patches within the same image to create challenging variants. It substantially increases effective dataset diversity and improves downstream performance.
Highlights
Produces hard, diverse training examples from a single image.
Preserves global semantics while perturbing local structure.
Complements standard flips, crops, color jitter, random grayscale, and blur methods.
Page updated
Google Sites
Report abuse