Planar Matchmove using Invariant Image Features

Sawhney et al. [1] perform automatic multi-view feature tracking in images from a moving camera to estimate 3D camera pose and scene structure. They then use this 3D location information to insert various graphics into the scene. Their paper is a bit vague about whether the feature tracking algorithm that they employ is novel, or borrowed, but they do admit to using a combination of strategies for feature tracking. They use multiscale patch based auto-correlation surfaces to identify features, and they compute the global camera pose and 3D geometry estimates using an iterative bundle block adjustment.