2D-to-Stereo Panorama Conversion Using GAN and Concentric Mosaics

Abstract

We describe a learning-based technique to automatically convert a 2-D panorama to its stereoscopic version. In particular, we train a generative adversarial network using perspective stereo pairs as inputs. Given a 2-D panorama, we partition it into overlapping local perspective views. To satisfy the panoramic stereo condition, we generate a sequence of left and right stereo view pairs and stitch them to produce concentric mosaics. We also describe experiments on synthetic and real datasets as well as comparisons with competing state-of-the-art techniques, which validate our technique.

Keywords

• 2D-to-stereo panorama
• GAN
• concentric mosaics
• depth peeling loss
• selector
• stereo panorama synthesis