Intra-Class Part Swapping

• @zhangIntraClassPartSwapping2021
• replaces most attentive regions of one image by the other
• Attentive regions are extracted using a classification activation map (CAM), thresholded for the most prominent region
• The attentive region in the source image is scaled and translated according to the attentive region of the target image for region replacement
• The label information of the output is similar to the target image as this approach relies on augmenting similar class images.