The method of Faugeras and Keriven for solving the multiple view stereo problem by partial differential equations (PDE's) in a level set implementation has been well received. The main reasons are that it produces good results and deals effectively with objects which are, topologically, more complex than a ball. This makes it a good choice for the final step in the usual structure from motion approach, namely making a full 3D surface reconstruction from estimates of camera orientations and 3D point structure. Here an approach is proposed whereby the 3D point structure, estimated using structure from motion, is used to initialize the method of Faugeras and Keriven. The proposed approach has the advantage of a) considerably improving the run time of the approach, and b) making it more resistant towards noisy data and data with patches of low variability. Both advantages make the approach much more effective on real data.