Deformable Template Based Stereo

We present a new deformable template based technique for recovering depth information from two or more images. The basic idea of our method is to use 3-D planar patches as the primitives describing the surfaces in a scene. In order to recover the depth information of a planar patch, two constraints on the view-line and the imaging geometry, and a deformable template matching based hypothesis-and-verification procedure are introduced. Each candidate of the planar patch is projected perspectively onto the observed images, and the image brightness or edge intensity differences between the projected images and the observed ones are calculated. The candidate which has the best fitting with the observed data will be selected as the solution. The proposed method can obtain a dense and accurate depth map directly from the input images.