Speeding up stereo algorithms has been an ongoing problem in computer vision. Much research has been done in developing specialized hardware that implements the stereo algorithm in parallel [7]. Our algorithm could exploit special hardware, but it is particularly well suited for implementation on sequential computers, which are currently much less expensive and easier to program.
Our approach is related to the coarse-to-fine stereo algorithm [15] which takes advantage of the results obtained at lower resolutions to predict the results at larger resolutions. This approach is related to our work because the coarse-to-fine algorithm is designed to take initial values for stereo search in order to speed up the computation. In the coarse-to-fine approach, information is propagated between resolutions, while our approach propagates information temporally.
The idea of temporally propagating knowledge about the scene in order to speed up the execution of an algorithm is not a novel one. The concept of temporal coherence is used in computer graphics to propagate the scene structure through time. For example, temporal and structural coherence is used in accelerating the calculation of animation sequences [5]. While the concepts used in graphics are similar, the overall goal is different. Computer graphics generates images given the scene structure, and the stereo algorithm produces depth maps given the images.
Recent work in view synthesis is also related to our work. View synthesis is concerned with generating realistic-looking images of a scene from a novel viewpoint, given one or more images of the scene. Work done by Scharstein [13] uses a stereo image pair to generate views from new viewpoints. Our approach is similar to this work because our algorithm needs the depth maps from new viewpoints. We share the problem of obtaining new disparity maps given sparse information. The difference, however, is in the application of the new depth maps. While view synthesis is concerned with reproducing the images from the new depth maps, we use the new depth maps to improve the performance of the stereo algorithms.
There has been much research in results of stereo algorithms over time [12], [1], [14]. The bulk of this research uses only results of the stereo algorithm, without altering its performance. To the best of our knowledge, the idea of temporally extending the stereo results in order to accelerate the algorithm is a novel one.