The point is simply to use the extra information that at silhouettes borders, "there is one point where the surface meets the line of sight and the surface normal is perpendicular". This is provides more constraints on the 3D shape than just taking the intersection of silhouettes. It helps recover a more accurate and regular surface.