Generating Mosaics with Minimum Distortions

Manifold mosaicing is a fast and robust way to summarize video sequences captured by a moving camera. It is also useful for rendering compelling 3D visualizations from a video without estimating the 3D structure of the scene. However, since the result mosaics are not perspective images, their geometry is inherently distorted. These mosaics are commonly referred to as multi-perspective images, or multi-perspective mosaics. In this paper we address the following question: Given a video captured by a moving camera, what is the best multi-perspective mosaic that can be generated from it? What is the mosaic with the best combination of large field-of-view and minimal geometric distortions? We define the necessary conditions for a good mosaic and a quantitive criterion for the geometric distortions, and derive analytically the optimal mosaic under this criterion. Results on video sequences confirm that indeed the optimal mosaic has significantly better quality than those generated by other techniques.