Dense Structure Estimation via Regularised Optical Flow

One of the major diAEculties in stereo applications is the establishment of correspondences at every pixel in order to compute dense depth maps. Instead of the classic two camera system we employ a special experimental setup with only one camera which is continuously moved from one stereo position to the other. The correspondence problem is solved by tracking each pixel over the acquired image sequence. From a di erential total least squares method we obtain an initial optical ow estimation. A subsequent regularisation method allows us to recover dense full ow taking the con dence from the TLS estimation into account. The resulting structure estimation algorithm is tested on two real examples.