A constraint satisfaction network for matching 3D objects

A new approach is presented for matching visible surfaces of 3D objects using a constraint satisfaction network. This in turn provides the necessary basis for volumetric reconstruction from multiple views. By matching, the authors mean both to establish correspondence between individual faces and to compute 3D transform that would bring one in correspondence with the other. Toward this goal, constraints at three different levels of complexities are specified to produce stable and coherent assignments. The constraint satisfaction is implemented as a Hopfield network with an appropriate energy functional and minimized using simulated annealing. The system extracts objects faces by computing their bounding contours with adaptive scale space filtering. This process identifies important surface features such as jumps or occluding boundaries and creases. The pointwise feature descriptors are then linked, and an attributed graph is derived to represent the object. The nodes in the graph represent geometric surface features, and the links in the graph represent the relationship between adjacent surfaces. The authors present results on real images.<>