Most automated 3D acquisition systems similar to the approach presented in the paper by Rusinkiewicz et al. [1] rely on structured light patterns projected onto the surface of the models, as well as human interaction in order to deliver acceptable results.

The authors of this paper describe a fully automated approach for textured 3D models using an active, well-calibrated stereo head. Compared to the space carving method, which demands a large number of images, the proposed approach requires only a few images and relies on stereovision techniques to reconstruct 3D models of acceptable accuracy. The approach is based on the asymptotic Bayesian method for reconstructing missing depth-values of a series of 2D images. In order to keep the processing overhead low, a voxel-based data structure called inverse polar octree (IPO) is used. The motion path of the stereo head is determined automatically, based on edge images of the camera views. The proposed approach is demonstrated by means of two different scenes, where five and 20 viewpoints, respectively, are used to reconstruct sufficiently accurate 3D models of two scenes.

With respect to the given examples, the approach seems to work very well. It should be noted that this approach is completely automated, and the motion planning of the stereo head is optimized in order to minimize the estimated error of the Bayesian method. Unfortunately, there are no runtime figures available.

The main contribution of this work is a model acquisition approach that uses less image samples than comparable space carving methods. It takes into account surface textures, and does not rely on projected light patterns. This is accomplished by using a calibrated binocular camera head instead of a monocular system. In addition, the reconstruction process focuses on areas of the current view where the stored voxel data is inconsistent with the images taken from the viewpoint of the binocular head.