Bootstrap based cooperative processes in computer vision

A new approach for executing computer vision tasks is presented. In real situations the complexity of the input data and/or computational procedure limits the possibilities of rigorous modeling, and therefore it is difficult to design algorithms optimal for a wide variety of operating conditions. Validating the assumptions embedded into a computer vision algorithm for the given input is a necessary condition if robust techniques are desired. We propose the use of bootstrap based cooperative processes for validation. The set of outputs, obtained by perturbing the input data in the execution of the task, defines the empirical distribution of the output. From the distribution an output confidence measure under the given operating conditions can then be assessed. Based on these confidence values the task can be executed using less constraining assumptions about the data and thus improving the robustness of any algorithm. The derived confidences also provide tools for evaluating the performance of the system under realistic operating conditions. The proposed approach is motivated by resampling techniques developed in statistics during the past decade, especially the bootstrap. The bootstrap method is a nonparametric estimation technique of the statistical behavior of an estimate when only a single sample of the input data is available. We make extensive use of bootstrap techniques. The methodology of using cooperative processes is first applied to evaluate and compare the performance of several edge detection systems. Confidences are obtained by using the perturbation of nuisance properties of the input, properties with no relevance for the output under ideal conditions. Based on the confidence values, an edgemap independent of the gradient magnitude is derived, As another example we show that robust image segmentation can be achieved based on the consensus information extracted from the output of several region-adjacency-graph (RAG) pyramids having a probabilistic component. The generality of the new technique is discussed and its applications for other computer vision tasks are proposed as further research.