A New Framework for Multi-camera Structure from Motion

Abstract

Models of real-world objects and actions for use in graphics, Virtual and augmented reality and related fields can only be obtained automatically through the use of visual data and particularly video. This paper introduces a new framework for integrating multiple synchronized videos of an object or a scene captured by a calibrated configuration of cameras into a unified space-time description. Given videos of a rigidly-moving object, it is a first step for dynamic descriptions and model building to recover the three-dimensional (3D) motion of the object which consists of a rotation and a translation at each instant. In this paper we extend the structure-from-motion framework for moving cameras in a static world to static multi-camera configurations consisting of large numbers of cameras observing a moving world. This new framework integrates the information from all the cameras into a single model. Using all the given motion constraints at once, it does not suffer from the intrinsic ambiguities of single-camera approaches. We show this by arguments based on the distribution of camera locations in space. Finally, we present a new algorithm that implements this new framework and computes the motion and structure of a rigidly moving object observed by 50 calibrated cameras.