A Bottom-up and Top-down Based Disparity Computation

It is becoming apparent that stereo matching algorithms need much information from high level cognitive processes. Otherwise, conventional algorithms based on bottom-up control alone are susceptible to local minima. We introduce a system that consists of two levels. A lower level, using a usual matching method, is based upon the local neighborhood and a second level, that can integrate the partial information, is aimed at contextual matching. Conceptually, the introduction of bottom-up and top-down feedback loop to the usual matching algorithm improves the overall performance. For this purpose, we model the image attributes using a Markov random field (MRF) and thereupon derive a maximum a posteriori (MAP) estimate. The energy equation, corresponding to the estimate, efficiently represents the natural constraints such as occlusion and the partial informations from the other levels. In addition to recognition, we derive a training method that can determine the system informations from the other levels. In addition to recognition, we derive a training method that can determine the system parameters automatically. As an experiment, we test the algorithms using random dot stereograms (RDS) as well as natural scenes. It is proven that the overall recognition error is drastically reduced by the introduction of contextual matching.