RICAP

• @takahashiDataAugmentationUsing2020
• Random image Cropping and patching
• Cropping four regions from randomly sampled images and augmenting them to create a new image.
• The generated image has mixed labels proportional to the pasted area
• The area of cropped regions in the output image is determined by sampling through uniform distribution
• anywhere-RICAP (origin can be anywhere), center-RICAP (origin can only be in the middle of the image), and corner-RICAP (origin can only be in corners
• Corner-RICAP has shown the best performance because a larger region of one image is visible to the network to learn