Synthesizing novel views for Street View experience

Abstract

Navigational applications often suffer from restricted and granular movement possibilities caused by a limited capture of real-world locations. Even the largest collections of street photos like Street-View, Mapillary [31], and SPED win more in geographical coverage than in qualitative capture of specific scenes. A possible solution to this problem could be post-processing of available image collections and generation of new photos that would restore the missing parts. This is the task of novel view synthesis - a known area in computer graphics and vision, that has shown impressive results over last several years [26], [27], [33], etc. However, the problem of real-world outdoor scene reconstruction is the most challenging, and is still a subject to active research. In this work we will explore different approaches to novel view synthesis and evaluate some of them on the sparse real-world imagery from Street-View dataset.