Metric reconstruction based on multifocal tensors

Reconstructing the structure of a scene from 2D views has been an important research field in computer vision. By investigating the epipolar geometry between multiple views, this paper deals with the problem of reconstructing metric 3D scenes from calibrated images and develops a uniform reconstructing framework based on multi-linear relations. The framework can reconstruct the scene directly and puts forward no requirements on the beforehand knowledge of camera movements and the scene´s 3D structure. First, a multifocal tensor is computed from the normalized images. Then, the tensor is applied to analyze the relative positions of views. Finally, local estimations are unified into an accordant entirety to get optimal approximations for the points in 3D scenes. Experiments with both synthetic and real image data show the validity of the framework in precision and efficiency.