Yiangos Georgiou, Melinos Averkiou, Tom Kelly and Evangelos Kalogerakis

The creation of high quality textures for immersive urban environments is a central component of the city modeling problem. Many recent pipelines capture or synthesize large quantities of city geometry using scanners or procedural modeling pipelines. Such geometry is intricate and realistic,however the generation of photo-realistic textures for such large scenes remains a problem - photo datasets are often panoramic and are challenging to re-target to new geometry. To address these issues we present an unsupervised neural architecture to generate photo-realistic textures for urban environments. Our Projective Urban Texturing (PUT)system iteratively re-targets textural style and detail from real-world panoramic images to unseen, unstructured urban meshes. The output is a texture atlas, applied onto the input 3D urban model geometry. PUT is conditioned on prior adjacent textures to ensure consistency between consecutively generated textures. We show results for several generated texture atlases, learned from different cities, and present quantitative evaluation of our outputs.