Reconstruction of the Pose of Uncalibrated Cameras via User-Generated Videos

Extraction of 3D geometry from hand-held unsteady uncalibrated cameras faces multiple difficulties: finding usable frames, feature-matching and unknown variable focal length to name three. We have built a prototype system to allow a user to spatially navigate playback viewpoints of an event of interest, using geometry automatically recovered from casually captured videos. The system, whose workings we present in this paper, necessarily estimates not only scene geometry, but also relative viewpoint position, overcoming the mentioned difficulties in the process. The only inputs required are video sequences from various viewpoints of a common scene, as are readily available online from sporting and music events. Our methods make no assumption of the synchronization of the input and do not require file metadata, instead exploiting the video to self-calibrate. The footage need only contain some camera rotation with little translation --- for hand-held event footage a likely occurrence.