A nonparametric statistical approach for stereo correspondence

This paper introduces a novel non-parametric statistical metric that can decide if the recovered set of parameters from a computer vision optimization process can actually be considered as a statistically significant solution. The level of significance can be used as a quality metric of the solution which makes it possible (i) to compare the solutions obtained using different optimization methods, and also (ii) to set intuitive thresholds on the acceptance criteria. We chose the stereo correspondence optimization methods as the initial test bed for the new technique. We compare and combine the results of sum of squared differences (SSD) and sum of absolute differences (SAD) methods for stereo correspondence. We validated our claims by running experiments on standard stereo pairs with ground truth. The results showed that the introduced ideas actually work very well and they can be used to improve the optimization results from different sources.