Puzzle Mix

• @kimPuzzleMixExploiting2020
• learns to augment two images optimally based on saliency.
• Images are divided into regions for the mixup
• The algorithm learns to transport the salient region of one image such that the output image has the maximized saliency from both images.
• $h(x_0,x_1)=(1-z)\odot {\Pi }_0^T{x}_0+z\odot {\Pi }_1^T{x}_1$
• where $z_i$ is a binary mask, $\lambda =\frac{1}{n}{\Sigma }_i{z}_i$ is the mixing ratio and ${\Pi }_0,{\Pi }_1$ are represent $n\times n$ grids that denote the amount of mass that is transported during transport of the image patch to another location.