GridMix is a data augmentation technique that divides images into a grid and randomly assigns each cell to one of two training images — creating a checkerboard-like mixing pattern that distributes information from both images evenly across the spatial dimensions.

How Does GridMix Work?

• Grid: Divide the image into an $n imes n$ grid of cells.
• Assignment: Randomly assign each cell to image $A$ or image $B$ with probability $lambda$.
• Mix: Fill each cell with the corresponding region from the assigned image.
• Labels: Mixed proportionally to the number of cells assigned to each image.

Why It Matters

• Spatial Distribution: Unlike CutMix (single contiguous region), GridMix distributes both images across the entire spatial extent.
• Multiple Regions: Forces the model to handle multiple disjoint regions from each class simultaneously.
• Complementary: Can be combined with other augmentation strategies.

GridMix is checkerboard image mixing — distributing both images across a grid for spatially diverse data augmentation.