Synchronization and calibration of camera networks from silhouettes

We propose an automatic approach to synchronize a network of uncalibrated and unsynchronized video cameras, and recover the complete calibration of all these cameras. In this paper, we extend recent work on computing the epipolar geometry from dynamic silhouettes, to deal with unsynchronized sequences and find the temporal offset between them. This is used to compute the fundamental matrices and the temporal offsets between many view-pairs in the network. Knowing the time-shifts between enough view-pairs allows us to robustly synchronize the whole network. The calibration of all the cameras is recovered from these fundamental matrices. The dynamic shape of the object can then be recovered using a visual-hull algorithm. Our method is especially useful for multi-camera shape-from-silhouette systems, as visual hulls can now be reconstructed without the need for a specific calibration session.