3D Reconstruction of Real Objects from silhouettes and stereo

Carlos Hernandez, Yucel Yemez, Francis Schmitt
Signal and Image Processing Department, ENST - CNRS URA820
46 rue Barrault, 75634 Paris Cedex 13, France

We present an automatic system for three dimensional reconstruction of real objects with high resolution shape and texture. This is a passive method working on a 2D image sequence delivered by a calibrated camera from different view angles, as the object rotates on a turntable. The surface reconstruction is obtained in two steps.

First the visual hull corresponding to the image sequence is extracted from the object silhouettes by volume carving. This step combines an octree construction with an adapted marching cubes algorithm. The resulting triangulated model defines the surface of a volume by excess containing the original object. To recover the hidden concavities we then deform in a second step the triangle mesh by sticking its vertices onto the object surface process by using stereo-techniques based on multi-correlation. Finally we use a texture mapping strategy based on surface particles which yields smooth texture transitions within the triangles and along their boundaries. Reconstruction results are shown to demonstrate the attained quality.