This project is about training GAN networks for image-to-image translation while keeping into account the entanglement effects due to occluders. You can find more information about our work in our paper.
In the specific case of rain, we use a realistic drop model that takes into account parameters related to drops size, shape, and probability to appear.
This is a demo that showcases our inference results, in which we trained a GAN only on rainy data, hence having both wetness and raindrops on the lens, and we were able to learn to render wetness only, while rendering raindrops with our physical model.