Perspective Preserving Style Transfer for Interior Portraits

Abstract

One of the jarring limitations of existing style transfer techniques is their failure to capture the illusion of depth through perspective. These often result in flat-looking images that have their style elements simply distributed across the image. Though recent methods attempt to alleviate this by considering depth information for a distinct stylization between foreground and background, they still fail to capture an image's perspective. When used on interior portraits where perspective is instinctively observed through its surfaces (walls, ceiling, floor), previous methods cause unwanted styling such as style elements distorting boundaries of surfaces and style elements not receding according to the perspective of the surfaces. In this paper, we developed a novel approach to effectively preserve interior portraits' perspective during style transfer, yielding stylized images that distribute and warps style elements according to the interior surfaces' perspective. Our method involves removing the perspective information from an interior portrait image, such that when performing style transfer the image can be considered as a flat perspective-neutral canvas. After that, we restore the perspective to the image leading to its style elements recede towards the vanishing point of its respective surface. We also observe that our approach was able to preserve depth information for some styles despite not extracting depth maps from the content.

Keywords

• Neural style transfer
• image processing
• image filters