Projective reconstruction recovers 3-D points in projective space P^3 from their positions in several 2-D images (P^2). Applications of projective reconstruction include algorithms for selecting point correspondences, algorithms for camera self-calibration and algorithms for 3D shape recovery.

Having m point correspondences across n images, the task of projective reconstruction leads to a system of m(choose 4 from 3n) polynomial equations of degree 4 in 11(n-1) unknowns.

At present, effective algorithms for projective reconstruction are known only from two, three or four views. The algorithms are based on multiple view tensors, i.e on the fundamenthal matrix, the trifocal or quadrifocal tensors. These methods decompose the nonlinear problem into two linear subproblems, the estimation of the tensor from image data and the subsequent decomposition of the tensor to projective camera matrices.

We introduce a method for the projective reconstruction from n views. The method is based on concatenation of trifocal constraints. The algorithm relies on linear estimates only. The method is not symmetrical with respect to input data. One of the captured images is selected as the reference image, which plays a special role during the computation. The proposed algorithm requires that all the points involved be visible in the reference image. The accuracy and stability of proposed algorithm with respect to pixel errors were tested. Experimental results will be presented.