Summer2Monsoon: Using CycleGAN for Image-to-Image Translation- 1 min
Image-to-image translation involves learning a mapping between in an input and an output image. Here we consider a particular case of converting images taken in a summer setting to monsoon setting and vice versa. This specific translation finds applications like training dataset augmentation for autonomous driving systems, video conversion in film industry where waiting all year long for a different season can be cumbersome process. We employ CycleGAN for this purpose, where two GANs learn the required mapping by achieving cyclic consistency. We conduct a perceptual study to quantify the translations obtained. Our method also provides a solution to the problem of single image de-raining and we compare our results to a recent work that attempts to do the same.
You can read more about the project here: LINK
Code: Our Summer and Monsoon dataset (~10K images) available upon request. CycleGAN repo LINK