Accurate Recovery of Dense Depth Map for 3D Motion Based Coding
Nikos Georgis, Josef Kittler, Mirek Bober · European Transactions on Telecommunications · 2000
Abstract The problem of scene structure recovery from image motion is considered in the context of motion compensated video coding. In order to compensate for full 3D motion of the camera, the scene depth map together with camera egomotion need to be estimated from the image sequence. An important prerequisite of the shape and motion recovery is the establishment of correspondences between the points in two successive frames of a video sequence. In the paper we abandon the traditional feature based approach and develop a correspondence analysis technique based on robust matching of local image intensities under affine transformation. We show that the technique not only establishes true correspondences, but the resulting field of corresponding points is dense, and their displacement is obtained with a subpixel accuracy and free of bias. We demonstrate that these dense high quality matches yield accurate estimates of the imaging geometry which are several orders of magnitude better than estimates obtained with feature based methods. This accuracy is reflected in a superior quality of estimates of the scene structure. The experiments are performed with images of a scene with known geometry, obtained using a calibrated camera. The obtained depth estimates are dense, with error of the order of 1% which contrasts with the sparse depth map and greater than 10% errors offered by a typical feature based approach. The technique is also applied to coding. Preliminary results indicate that significant improvement of the predicted image can be achieved by using accurate dense depth cstimates.