Semi-Supervised Image Deraining Using Gaussian Processes

Recent CNN-based methods for image deraining have achieved excellent performance in terms of reconstruction error as well visual quality. However, these are limited the sense that they can be trained only on fully labeled data. Due to various challenges obtaining real world fully-labeled datasets, existing synthetically generated data and hence, generalize poorly real-world images. The use training networks is relatively less explored literature. We propose a Gaussian Process-based semi-supervised learning framework which enables network derain using synthetic dataset while generalizing better unlabeled More specifically, we model latent space vectors Processes, then used compute pseudo-ground-truth supervising pseudo ground-truth further supervise at intermediate level Through extensive experiments ablations several challenging datasets (such Rain800, Rain200L DDN-SIRR), show proposed method able effectively leverage thereby resulting significantly compared labeled-only training. Additionally, demonstrate images GP-based results superior methods. Code available at: https://github.com/rajeevyasarla/Syn2Real.