In this paper, we present a novel method to fuse observations from an inertial measurement unit (IMU) and visual sensors, such that initial conditions of the inertial integration, including gravity estimation, can be recovered quickly and in a lin