Peter Hall
Department of Computer Science, University of1 Wales Cardiff
Cardiff CF2 3XF
peter@cs.cf.ac.uk
We present an algorithm for recovering a three-dimensional space-curve, rigid and open-ended, given a pair of images at arbitrary viewing angles. Our algorithm relies only on epipolar geometry, and connectivity of the curve. It is successful, even in the case of curves which appear self-intersecting, and is robust to noise. Only one end of the curve need be visible in either image, and as many solutions as the data and assumptions support are produced. Stronger assumptions can be built in as desired. Tests on control models show the algorithm is capable of sub-pixel accuracy, even in the presence of heavy noise. Tests on real data show it capable of reconstructing from a wider range of images than previous algorithms allow; we show it works where all other algorithms we tried failed.