Assessing generalization of pix2pix style transfer GAN for semantic segmentation of PV panels from aerial imagery
We don't have enough training datasets from developing countries. Models that generalize better can be more helpful in global development.
Generalization of deep learning models are found to have improved because of using GAN generated images as training images. Using a Conditional GAN may improve generalization performance.
We use Pix2Pix, a supervised learning CGAN that can be used for semantic segmentation by framing it as a style transfer problem. We assess the generalization performance of Pix2Pix.
Available in a Google Drive here