K.U.Leuven > ESAT > PSI > Visics > Research > Topics > Item 1.2

Registration of 3D surfaces through invariant feature matching

L. Van Gool, J. Vanden Wyngaerd

3D shape models are often put together from several partial reconstructions. Automatic registration can be seen as solving a three dimensional puzzle. As input, the algorithm receives a number of partial reconstructions. The goal is to create a single model by integrating all these parts. To do this, three steps are done:

1. Crude registration: positioning of the parts in single coordinate system.
2. Fine registration: refining the relative positions.
3. Integrating the parts into a single model.

There are good algorithms available now to perform the necessary, fine registration automatically, but only after the partial reconstructions have been brought into approximate positions. Research here focuses on techniques to do precisely that. Automatic "crude registration" is an important missing part for realizing completely automatic 3D recording systems.

The technique developed in our lab is based on the extraction of special curves on the surface, i.e., bitangent curve pairs. These bitangent curve pairs are characterized by invariant signatures. The invariant signatures are independent of the position of the surface in space and are used for finding corresponding points on different parts. The signatures can be found and matched efficiently. Registration of these curves yields the crude initialisation that more precise registration algorithms like ICP or mutual information maximization can start from.

The figure below shows an example of the results for 14 parts of a sculpture. On the left hand side, you see the input patches with the bitangent curve pairs. On the right hand side, you see the positioning of these parts. A movie illustrating the 3D registration process is available as reg3d.mpg.