Estimation of Camera Positions over Long Image Sequences

In this paper, we present an iterative algorithm that computes the camera path of long image sequences. It consists in applying successive bundle adjustment phases on different segments of the image sequence. The local models thus obtained are merged together into a common reference frame. The procedure is then repeated on a new grouping of the cameras, until the reconstruction error has reached a given error tolerance. The main objective is to ensure the scalability of the reconstruction and the good convergence of the bundle adjustment process by imposing a limit on the number of views for which the structure and motion parameters have to be simultaneously optimized. Error accumulation is also prevented by exploiting the presence of loopbacks and intersections in the camera path. We show results obtained over different camera paths, in particular a spiral path and snake-like path.