PhotoPath: single image path depictions from multiple photographs

Abstract

In this paper we present PhotoPath, an image synthesis technique that allows the visualization of paths through real world environments in a single intuitive picture. Given a series of input photographs taken with a hand-held camera along the path, our method generates an image that creates the impression of looking along the path all the way to the destination regardless of any curves or corners in the route. Rendered from a pedestrian perspective, this visualization supports intuitive wayfinding and orientation while avoiding occlusion of path segments in the image. Our approach intentionally avoids an involved and error-prone scene reconstruction. Instead we advocate the use of planar geometry proxies estimated from the sparse point-cloud obtained with Structure-from-Motion. The proxy geometry is spatially deformed in order to align the curved path with the straight viewing direction of the synthesized view. Finally, we propose a novel image composition algorithm accommodating for parallax and occlusion artifacts due to the approximation errors of the actual scene geometry.