Unsupervised Stereoscopic Video Style Transfer

Abstract

The creative style transfer across photos and videos involves transferring one style to another while maintaining the recipient image/video content. However, it is a challenging task, particularly for stereoscopic video, to preserve stereoscopic properties. This paper proposes a stereoscopic video style transfer method that maintains the temporal, depth, and stylization features. We suggest loosening the objective function to resolve the conflict between style transfer and temporal consistency for each left and right view to make the stylization loss term more motion-resistant. We use a zero-shot video style transfer framework for each left and right video frame. To make the stylized features consistent against the stereo features and consider the cross-view information in the stylized stereo video, we extend the parallax attention mechanism (PAM) as ePAM and use it to combine the left and right information. We compare quantitative and qualitatively with other image and video style transfer methods. Experimental results demonstrate the competitive performance of our method over the state-of-the-art. © 2023 Elsevier B.V., All rights reserved.