This paper investigates the problem of how to carry out 3D projective reconstruction from multiple images. Up to now

it is commonly accepted that the bundle adjustment and factorization method are the main methods for projective reconstruction. But the bundle adjustment needs a good initialization and extremely expensive computation

and the factorization method is limited by the restriction that all 3D points must be visible in all views. Recently

a linear algorithm of projective reconstruction based on the homography induced by the infinite plane was given by Hartley and Rother et al.

but they needed 4 points on a reference plane be visible in all views. This paper improves their algorithms and proposes a new linear algorithm based on infinite homography where 3 points on a reference plane should be visible in all views. It avoids the difficult task of determining whether 4 object points are coplanar or not

because 3 points which are not collinear just determines a plane. The algorithm proposed in this paper is convenient and can deal with the occluded problem. The algorithms are evaluated on both synthetic and real world images and the experiment results show that the method is accurate and only affected slightly by noise.