SHUNIT: Style Harmonization for Unpaired Image-to-Image Translation

We propose a novel solution for unpaired image-to-image (I2I) translation. To translate complex images with wide range of objects to different domain, recent approaches often use the object annotations perform per-class source-to-target style mapping. However, there remains point us exploit in I2I. An each class consists multiple components, and all sub-object components have characteristics. For example, car CAR body, tires, windows head tail lamps, etc., they should be handled separately realistic I2I The simplest problem will more detailed component than simple annotations, but it is not possible. key idea this paper bypass by leveraging original input image because include information about characteristics components. Specifically, pixel, we only gap between source target domains also pixel’s determine pixel. end, present Style Harmonization translation (SHUNIT). Our SHUNIT generates new harmonizing domain retrieved from memory an style. Instead direct mapping, aim styles harmonization. validate our method extensive experiments achieve state-of-the-art performance on latest benchmark sets. code available online: https://github.com/bluejangbaljang/SHUNIT.