Core Concepts
RoNet models the continuous generation of images across domains by learning an in-plane rotation over the style representation, achieving smooth and realistic transitions without requiring both ends of the translation line.
Abstract
The paper proposes a novel rotation-oriented solution, RoNet, to achieve continuous image-to-image (I2I) translation. Unlike typical linear interpolation approaches, RoNet models the continuous generation by learning an in-plane rotation over the style representation of an image.
Key highlights:
RoNet implants a rotation module in the generation network to automatically learn the proper rotation plane while disentangling the content and style of an image.
To encourage realistic texture, especially for challenging scenes like forests, RoNet designs a patch-based semantic style loss that learns the different styles of similar objects across domains.
Experiments on various tasks, including season shifting, real face to comic portrait, solar day shifting, and iPhone to DSLR, demonstrate RoNet's superiority in generating smooth and continuous translation results with a single input image.
RoNet outperforms state-of-the-art methods in both qualitative and quantitative evaluations, producing the most realistic and continuous translation results.
Stats
The paper does not provide any specific numerical data or statistics to support the key claims.
Quotes
"To achieve continuous I2I translation, we propose a novel rotation-oriented mechanism which embeds the style representation into a plane and utilizes the rotated representation to guide the generation."
"To produce realistic visual effects on challenging textures like trees in forests, we design a patch-based semantic style loss. It first matches the patches from different domains and then learns the style difference with high pertinency."