I found this amazing new research paper "High-Resolution Photorealistic Image Translation in Real-Time: A Laplacian Pyramid Translation Network" and an associated video.
It's about a new fast method for image-to-image translation. I wonder if there is any possibility to use such approach (separation of low-frequency and high-frequency components) for other kind of tasks. Like image classification, localization, segmentation and so on. We could process smaller, low-resolution images and the high-resolution details separately. And then we could merge the results. And at the end, instead of upsampling and creating new image, we could just add MLP and do the classification task.
I would really appreciate your answer along with the explanation why it is great, or why it doesn't make sense.