Shape refinement for reconstructing 3D-objects using an analysis-synthesis approach

We propose a novel shape refinement method for reconstructing the shape of a rigid 3D-object using multiple calibrated camera images. The presented method can be divided into two steps. Firstly, an initial shape model of the object is reconstructed using the well known shape-from-silhouette approach. Since silhouettes do not provide any information on object concavities, in a second step the initial shape model, represented by a triangular mesh, is iteratively refined using an analysis-synthesis approach. Therefore, in each iteration cycle local deformations are applied by shifting vertices of the triangular mesh. These local deformations are evaluated trianglewise by mapping texture from a respective reference input image to the current surface triangle, projecting the triangle to the image planes of all other cameras that observe it and comparing the obtained luminance distribution of the model views with the original input images. Minimising the differences between synthetic model views and input images yields a maximum photo-consistent model. The proposed method has been successfully tested on several real objects, example results obtained for two of these objects are given