Research Abstracts - 2006

Projective Surface Matching of Colored 3D Scans

Kari Pulli, Simo Piiroinen, Tom Duchamp & Werner Stuetzle

When scanning a 3D object surface, the individual scans are represented in the separate coordinate systems. In order to construct a coherent surface description, the scans can be registered so that when they are expressed in a single coordinate system the matching parts of the surface align well. Most registration methods align the data using geometry alone, or use the color information to obtain geometric registration constraints, still using only geometry information for the final alignment.

In this paper we present an approach that equally uses both geometry and color information. An initial registration is iteratively refined by minimizing an error function. The error function is defined by projecting the colored scan data into imaginary image planes of each scan location, and comparing discrepances both in range and color domains.

Colored 3D scans are projected onto the image planes of two virtual cameras. The registration is iteratively improved so that both colors and ranges align on the image planes. The paper derives the derivatives of the registration error with respect to registration transformations and uses them to efficiently obtain a good registration.

At first, the projection of right bunny scan does not align well with that of the left bunny (lower left image), nor does the left projected bunny align with the right one (lower right image).

Scans move in 3D so their projections align better...

... and finally the scans align as well as they can.
(On the left, the "right" scan is projected over the silhouette of the "left" scan, on the right it is vice versa). 

More information can be found in the paper.